Proceedings

Collaborative team recommendation involves selecting users with certain skills to form a team who will, more likely than not, accomplish a complex task successfully. To automate the traditionally tedious and error-prone manual process of team formation, researchers from several scientific spheres have proposed methods to tackle the problem. In this tutorial, while providing a taxonomy of team recommendation works based on their algorithmic approaches to model skilled users in collaborative teams, we perform a comprehensive and hands-on study of the graph-based approaches that comprise the mainstream in this field, then cover the neural team recommenders as the cutting-edge class of approaches. Further, we provide unifying definitions, formulations, and evaluation schema. Last, we introduce details of training strategies, benchmarking datasets, and open-source tools, along with directions for future works.

Conversational User Interfaces (CUIs) have been argued to have advantages over traditional GUIs due to having a more human-like interaction. The growing popularity of conversational agents has enabled humans to interact with machines more naturally. People are increasingly familiar with conversational interactions mediated by technology due to the widespread use of mobile devices and messaging services and a hungry market for conversational agents. Based on the recent advances in conversational AI, due to the proliferation of large language models, there are clear signs that the future of human-computer interaction will have a significant conversational component. Today, over two-thirds of the population on our planet has access to the Internet, with ever-lowering barriers to accessibility. This tutorial will showcase the benefits of employing novel conversational interfaces for crowd computing, human-AI decision making, health and well-being, and information retrieval. Given the widespread adoption of AI systems across several domains, we will discuss the potential of conversational interfaces in facilitating and mediating people’s interactions with AI systems and the opportunities and challenges that lie at this intersection from the user modeling and personalization standpoint. The tutorial will include interactive elements and discussions and provide participants with practical insights to inform the design of effective conversational interfaces.

The objective of this tutorial is to provide the researchers of the UMAP community with methodologies, tools and techniques to model complex psychomotor behaviours that can later personalize learning support in realms like sports, physical education or for rehabilitation purposes, providing insights into data gathering from activities that involve human movement. Research in the psychomotor field to provide personalization support to users poses several research challenges. In this tutorial we will focus on how to model psychomotor learning by exploring the fusion of computational analysis of the learning process and wearable / video analysis. Participants will engage in hands-on activities, recording specific movements and learning how to capture human body keypoints. Throughout the training, various tools are introduced, allowing participants to explore options that align with their specific needs and expertise. This tutorial aims to contribute to the thriving initiatives within the UMAP community, fostering the modeling of physical activities.

The presented tutorial aims to serve as a comprehensive roadmap for the UMAP community into the current user modeling research, focusing on the paradigm shifts that have transformed the research landscape in recent times. We will provide a complete overview of the large, long-standing, and ever-growing research fields of user modeling and user profiling, both from a historical and a technical point of view. We will then examine the definitions associated with each key term in this research domain, aiming to eliminate ambiguity and confusion in their usage. As the core of our tutorial, we present in-depth the paradigm shifts that have occurred in recent years, especially due to technological evolution, as well as the current research directions and novel trends in the field. In particular, we illustrate and discuss the advances in the following topics: implicit and explicit user profiling, user behavior modeling, user representation, and beyond-accuracy perspectives. The audience will be engaged in discussions during the whole presentation to foster the development of an interactive event.

Detailed information and resources about the tutorial are available on the website: https://link.erasmopurif.com/tutorial-umap24.

This tutorial provides an interdisciplinary overview of fairness, non-discrimination, transparency, privacy, and security in the context of recommender systems. According to European policies, these are essential dimensions of trustworthy AI systems but also extend to the global debate on regulating AI technology. Since the aspects mentioned earlier require more than technical considerations, we discuss these topics from ethical, legal, and regulatory perspectives. While the tutorial’s primary focus is on presenting technical solutions that address the mentioned topics of trustworthiness, it also equips the primarily technical audience of UMAP with the necessary understanding of the social and ethical implications of their research and development and recent ethical guidelines and regulatory frameworks.

Teachers report difficulties engaging in search in the context of lesson planning. Despite the availability of customised search solutions like Learning Object Repositories, most teachers report using general-purpose search engines like Google. Unlike students – to whom most research in educational search is directed – teachers are experts capable of using many types of resources for their tasks. In addition, since they teach multiple classes, they switch educational contexts frequently. These circumstances present unique challenges for user modelling. This thesis investigates how an adaptive browser-based search engine augmentation could support teachers in both exploratory and lookup search tasks. A preliminary user study was conducted to understand the actual search behaviour of teachers in their everyday environment and to investigate the inference of search tasks and context from search behaviour. Using this inferred information, personalized search aids could be delivered through the extension such as search engine result page (SERP) summarization for exploratory search and context-based search snippets. The relevance of existing search snippets was investigated and potential for improvement using teachers’ search context was identified. With the rest of the PhD, I plan to expand the user study to get more concrete results about teacher search behaviour and implement and test this adaptive browser extension to see how it affects the search experience for teachers.

Recommender systems have become essential in aiding users’ decision-making processes, yet ensuring users understand the rationale behind recommendations remains a challenge. Balancing the presentation of both positive and negative aspects of recommendations is important to foster user trust and satisfaction. In this PhD thesis, I want to investigate the balance between supporting recommendations and alerting users to potential risks. While prior research has broadly examined persuasive and informative explanations, my focus is on exploring methodologies to effectively caution users about associated risks. By prioritizing the warning aspect, the aim is to heighten user awareness of biases and potential harms, ultimately enhancing transparency and trust in the recommendation process.

The flourishing presence of Artificial Intelligence (AI) in daily life emphasizes the necessity for education about AI. It is crucial to create tailored educational materials and tools to meet the unique educational needs of diverse groups. This ongoing project aims to develop personalised educational tools for AI literacy for different groups of society (i.e., children, teachers, and adults). The perception and knowledge of AI by these groups will be explored. Participatory design workshops with teachers and adults will assist in the development of personalised educational tools for these groups tailored to the specific needs of each group and their perceptions of AI. Finally, this project aims to identify the best practices for developing personalised educational tools for AI literacy for each group.

Item representations and item-to-item similarity measures are powerful tools used by many modern recommender systems (RS). Typically, recommender systems consider a singular representation of items and similarity measure between items. Studies have shown, however, that users can disagree on item descriptors and item similarities. If we consider that each user has their own subjective view of items, a singular item representation cannot cater to all users. We link this observation to recent studies showing that modern recommender systems approaches produce better recommendations for mainstream users, users whose preferences align with the majority/popular view, and poor recommendations for non-mainstream users with uncommon or niche tastes. It is our assumption that this bias towards mainstream users is due to the fact that (i) common RS datasets are dominated by mainstream users and (ii) if we again consider that each user has their own subjective view of items, an item representation built from a mainstream user dominated dataset will better represent the views of mainstream users and misalign with the subjective views of non-mainstream users. In this PhD work, we aim to explore methods which improve recommendations for non-mainstream users, by aligning the system’s representation of items with the target user’s subjective view of items during the recommendation step. Specifically, we focus on subjective perceptions of item distances and item characteristics.

Recommender systems offer an important technology aiding users in discovering relevant items among a multitude of available options. The emergence of Large Language Models (LLMs) enables a powerful opportunity to improve the performance and flexibility of recommender systems in different aspects. Those involve possibilities to enhance the item retrieval and ranking, explanation of recommendations, and generation of content to be recommended. The objective of my Ph.D. is to investigate those possibilities to identify and develop methods for improving different aspects of recommender systems using LLMs. The findings are expected to further develop the field of recommender systems and reveal novel opportunities for future research.

The fashion industry is a significant contributor to global carbon emissions, water consumption, and waste generation. My Ph.D. project explores two approaches to promote sustainability in fashion recommender systems. First, I aim to develop algorithms and user interfaces that nudge users towards more sustainable and ethical fashion choices by incorporating relevant data into the recommendation process and item presentation. This involves balancing user preferences with sustainability considerations and exploring effective ways to present this complex information to users. Second, I seek to make the recommendation algorithms themselves more environmentally sustainable by reducing their computational cost without compromising recommendation quality. To achieve these goals, I plan to leverage multimodal item representations, multi-criteria recommendation techniques, and novel algorithmic approaches. I expect to complete my Ph.D. in October, 2025.

Promoting mental health in adolescents is crucial to minimize the development of mental health disorders, especially for those with a mild intellectual disability who face more mental health challenges. In general, they have more difficulty controlling their feelings and show an overrepresentation in feelings of loneliness, anxiety, and other emotional problems. To improve their mental health, traditional teaching methods are unsuitable. Digital mental health intervention is a promising approach to do this in a fun, rewarding, and personally relevant manner. This PhD research uses a participatory approach to (1) identify the needs of these adolescents and (2) how to design, develop, and evaluate a digital intervention for mental health promotion. First school observations and interviews with staff suggest to improve self-reliance skills, finding help, and reflecting on wellbeing. The design process will focus on developing a personalized intervention that suits each adolescent’s needs, as home context and support network vary.

In recent years, as a result of the emergence of innovative technologies and applications, several solutions aiming to improve the quality of education have been appearing. However, mainly due to the human factor of lack of understanding and trust in these applications, adoption by higher education institutions remained low. In my doctoral thesis, I aim to contribute to overcoming this barrier in two main steps. First, I try to understand more deeply the nature of the existing contrast between perceived benefits and concerns of higher education students regarding one specific application area, artificial intelligence-based personalized learning. Building on these results, I will then design and evaluate a new personalized learning tool, developed specifically with user-centered data privacy and ethical considerations in mind that specifically respond to the identified concerns. The uniqueness of this tool is the transparency and controllability of different levels of personalization, which ensures that each student can freely choose the extent to which they are willing to make their data available in order to receive a personalized learning experience. The effects of this design principle on privacy concerns and learning outcomes will then be tested in multiple lab- and field studies. I believe this project fits well into both the “Personalizing Learning Experiences through User Modeling” and the “Fairness, Transparency, Accountability, and Privacy” tracks of the UMAP conference, which is why I have decided to apply for the Doctoral Consortium.

Empirical studies have extensively investigated human decision-making processes in various domains where AI systems are incorporated. However, comparing and replicating these studies can be challenging due to different experimental configurations. Moreover, the existing contexts often have limited scope and may not fully capture the complexity of real-world decision-making scenarios that are riddled with varying levels of uncertainty. Our framework addresses these practical gaps by providing a configurable and reproducible environment for conducting human-AI decision-making studies in the route planning domain that captures many complexities of real-world scenarios. Researchers can customize parameters, conditions, and factors involved in decision-making tasks to help address research and empirical gaps through rigorous experiments. With various modules such as map generation, chat components, and different AI systems available within the “DecisionTime” framework, researchers can effortlessly design experiments exploring multiple aspects of human-AI interaction and decision-making.

Personality has been introduced in recommender systems to address the cold-start problem, or to improve the recommendations by personalizing the degree of diversity for the specific user. For group recommender systems, personality has also been considered to select the most appropriate strategy for aggregating individual preferences or to model the dynamics of the group decision-making process. The most widely used models in this context are the Five Factor Model (FFM) and the Conflict Resolution Styles model (TKI or ROCI II). However, the challenge of eliciting this information remains an open problem, along with establishing correlations between self- and peer-evaluations of personality. In this work, we present the results of a preliminary study investigating the correlations between self and peer evaluations of personality. Furthermore, we analysed the correlations between the two personality models. Our findings show good consistency between self and peer personality evaluations for the FFM, but not for conflict resolution styles. This suggests that the FFM peer evaluations could be considered an alternative to self-evaluations. In contrast, for the conflict resolution styles, it may be necessary to measure both self and peer evaluations. This is relevant in group recommendation scenarios, where how group members perceive each other could impact the group decision-making process. Finally, our findings reveal correlations between the two personality models but do not support a consistent way to derive one from the other.

With the increasing adoption of Artificial Intelligence (AI) systems in high-stake domains, such as healthcare, effective collaboration between domain experts and AI is imperative. To facilitate effective collaboration between domain experts and AI systems, we introduce an Explanatory Model Steering system that allows domain experts to steer prediction models using their domain knowledge. The system includes an explanation dashboard that combines different types of data-centric and model-centric explanations and allows prediction models to be steered through manual and automated data configuration approaches. It allows domain experts to apply their prior knowledge for configuring the underlying training data and refining prediction models. Additionally, our model steering system has been evaluated for a healthcare-focused scenario with 174 healthcare experts through three extensive user studies. Our findings highlight the importance of involving domain experts during model steering, ultimately leading to improved human-AI collaboration.

Recommender systems play a crucial role in social media platforms, especially in the context of news, by assisting users in discovering relevant news. However, these systems can inadvertently contribute to increased personalization, and the formation of filter bubbles and echo chambers, thereby aiding in the propagation of fake news or misinformation. This study specifically focuses on examining the tradeoffs between the diversity of news recommendations and the dissemination of misinformation on social media. We evaluated classical recommender algorithms on two Twitter (now X) datasets to assess the diversity of top-10 recommendation lists and simulated the propagation of recommended misinformation within the user network to analyze the impact of diversity on misinformation spread. The research findings indicate that an increase in news recommendation diversity indeed contributes to mitigating the propagation of misinformation. Additionally, collaborative and content-based recommender systems provide more diversity in comparison to popularity and network-based systems, resulting in less misinformation propagation. Our study underscores the crucial role of diversity recommendations in mitigating misinformation propagation, offering valuable insights for designing misinformation-aware recommender systems and diversity-based misinformation intervention.

Calibration in recommender systems is an important performance criterion that ensures consistency between the distribution of user preference categories and that of recommendations generated by the system. Standard methods for mitigating miscalibration typically assume that user preference profiles are static, and they measure calibration relative to the full history of user’s interactions, including possibly outdated and stale preference categories. We conjecture that this approach can lead to recommendations that, while appearing calibrated, in fact, distort users’ true preferences. In this paper, we conduct a preliminary investigation of recommendation calibration at a more granular level, taking into account evolving user preferences. By analyzing differently sized training time windows from the most recent interactions to the oldest, we identify the most relevant segment of user’s preferences that optimizes the calibration metric. We perform an exploratory analysis with datasets from different domains with distinctive user-interaction characteristics. We demonstrate how the evolving nature of user preferences affects recommendation calibration, and how this effect is manifested differently depending on the characteristics of the data in a given domain. Datasets, codes, and more detailed experimental results are available at: https://github.com/nicolelin13/DynamicCalibrationUMAP.

Parkinson’s disease imposes a substantial global health burden and causes severe motor impairments that compromise the quality of life. Virtual reality-based mirror training (VRMT) has the potential to improve motor function through immersive action observation, motor imagery, and real-time feedback. This study aims to propose a novel VRMT personalized for Parkinson rehabilitation and present reference learning data from a biomechanical model obtained from twelve healthy individuals performing the activities of lifting the leg, arising from a chair, and touching the nose included in the Movement Disorder Society Unified Parkinson’s Disease Rating Scale. The results reveal insights on the maintenance of balance when arising from a chair through linear variation of center of mass, minimal hand position shaking and thus minimal tremors during activity of touching the nose, and high maximum foot velocity of approximately 0.79±0.26 1/s during activity of lifting the leg. The data collected accurately represents the expected healthy execution and related variability for each activity, serving as a learning reference for the future development of VRMT.

Comp-HuSim (Complex Human Simulations) is a multi-agent persona simulation platform powered by generative AI models (GPT-3.5 turbo, GPT 4 turbo, and Mixtral(8×7B)). Each Comp-HuSim generative agent is created with random variables such as age, gender identity, economic status, Big-5 personality characteristics, geographic location, etc. Agents are then imbued with backstory memories, textured histories, diverse hobbies and interests, and a series of distinct attributes. In early experiments our agents engage in conversation with one another, reflect on their past experiences, and play games together in surprisingly sophisticated ways. Complex forms of human-like behaviors become visible in the Comp-HuSim project demonstration, demonstrating our approach’s promise for applications in Virtual Assistants, Conversational Interactions, and Personalized Human-Robot Interaction.

Mixed Reality (MR) technologies have transformed the way in which we interact and engage with digital content, offering immersive experiences that blend the physical and virtual worlds. Over the past years, there has been increasing interest in employing Artificial Intelligence (AI) technologies to improve user experience and trustworthiness in cultural contexts. However, the integration of Large Language Models (LLMs) into MR applications within the Cultural Heritage (CH) domain is relatively underexplored. In this work, we present an investigation into the integration of LLMs within MR environments, focusing on the context of virtual art exhibitions. We implemented a HoloLens MR application, which enables users to explore artworks while interacting with an LLM through voice. To evaluate the user experience and perceived trustworthiness of individuals engaging with an LLM-based virtual art guide, we adopted a between-subject study design, in which participants were randomly assigned to either the LLM-based version or a control group using conventional interaction methods. The LLM-based version allows users to pose inquiries about the artwork displayed, ranging from details about the creator to information about the artwork’s origin and historical significance. This paper presents the technical aspects of integrating LLMs within MR applications and evaluates the user experience and perceived trustworthiness of this approach in enhancing the exploration of virtual art exhibitions. Results of an initial evaluation provide evidence about the positive aspect of integrating LLMs in MR applications. Findings of this work contribute to the advancement of MR technologies for the development of future interactive personalized art experiences.

Understanding the characteristics of the space of available items is instrumental for successfully exploring the options in interactive recommender system approaches, including conversational recommender systems. Forming an adequate mental model of the distribution of items with respect to their various features can prevent mismatches with the user’s preferences, either represented in an existing preference model or, even more importantly, in a process of preference development during the interaction. However, such a mental model is limited due to complex and opaque feature dependencies manifested in the item space that make it hard to predict whether a combination of feature preferences is realistically satisfiable or should be partially relaxed. Illuminating these dependencies in the recommender interface while the user scrutinizes their preference model has the potential to convey richer structural information, facilitate mental model formation, and thus allow better-informed decisions for both preferred feature ranges and recommended items. To this end, we introduce novel dynamic ridge plot sliders that employ guidelines from research on information visualization for mental model formation. The proposed sliders relate feature-wise item distributions to items’ overall utility values and in doing so provide predictive power about the implications of future user actions and preference adjustments. Furthermore, paradigms for interactions between multiple such sliders and other GUI elements are motivated via an interactive recommender prototype. We also present a preliminary user study with yet inconclusive results, and discuss the future potentials and conceivable application areas of our approach, as well as promising next research steps.

The recent boost in generative artificial intelligence has also reached the field of recommender systems. However, as is often the case, much of the work focuses on the algorithms, overlooking the crucial aspect of improving the systems from a user perspective. In this initial research, we explore the potential of large language models to achieve improvements in preference elicitation. The interactive choice-based method we are augmenting has previously demonstrated significant improvements in a number of aspects related to the user experience. Through an exploratory user study, we show that the item set comparisons presented by this method can be successfully accompanied by independently generated textual summaries, thereby improving the user experience even further.

Social robots are emerging as a valuable tool in promoting holistic person-centered care and well-being among seniors. One aspect that often affects seniors’ lives is the feeling of loneliness and social isolation. In the context of the SISTER project, we aim to investigate the acceptance of a personal social robot as a tool to mitigate these aspects. Empathy plays an important role in establishing a human-robot emotional and social connection. As a first phase of our project, we investigate empathy’s impact on human-robot interaction when a social robot serves as a senior companion. For this purpose, two robot dialog versions were developed: a non-empathic version, acting as a conversational interface without considering users’ emotional states, and an empathic version in which emotion recognition is used to understand and respond empathically to the users’ emotional state. Thirty older adults were involved in the study, revealing that, as far as senior-robot interaction is concerned, the empathic condition resulted in a more usable and positive experience showing that this robot’s capability should be considered in future developments for personalizing the interaction.

Mining detailed content from mobile human-computer interactions often relies on broad log files, which limits the specificity of analyses. Our study presents a unique approach capitalizing on Optical Character Recognition to continuously detect keywords across all applications and media formats, correlating this with system logs for context and duration tracking. Such a strategy even allows tracing of topics within pictorial content like memes on social media. A privacy concept based on a whitelist for keywords and topics and anonymized log files addresses typical concerns of potential study participants regarding their personal data. In our four-month study involving 25 participants, we generated an expansive nine-million-point dataset. This detailed dataset not only validates the efficacy of our approach but also exemplifies its capacity for rich, cross-app, long-term interaction analysis. In order to clarify data protection issues, we conducted qualitative interviews with eight of the participants on a voluntary basis.

Paper and digital forms are widely used to collect user information across multiple domains, such as research, healthcare, and education. However, both types still lack application and follow-up interpretation: Paper forms need to be digitised meticulously to be analyzed or shared with team members efficiently; in comparison, digital forms cannot convey handwriting and might require technical literacy. We present the FormTwin data collection tool—an alternative to online forms, which allows for the efficient reuse of existent paper forms while providing the convenience of digital forms. FormTwin can digitise a wide range of forms with the integrated form annotation application. Then, it combines two input channels: A stylus on a tablet and a digital smart pen on plain paper, which duplicates the input on a mobile application in real-time. We aim to improve the efficiency and accessibility of data collection for practitioners with a modular system that combines both the digital accessibility of digital forms with keeping the needed technical literacy low and retaining the quality of hand-drawn sketches.

In recent years, robotics has evolved, placing robots in social contexts, and giving rise to Human-Robot Interaction (HRI). HRI aims to improve user satisfaction by designing autonomous social robots with user modeling functionalities storing data on people to personalise interactions. Personality, a vital factor in human interactions, influences temperament, social preferences, and cognitive abilities. Despite much research on personality traits influencing HRI, little attention has been paid to the influence of the robot’s personality on the user modeling. Personality can influence not only temperament and how people interact with each other but also what they remember about an interaction. A robot’s personality traits could therefore influence what it remembers about the user and thus modify the user model and the consequent interactions. However, no studies investigating such conditioning have been found. This paper addresses this gap by proposing distinct user models that reflect unique robotic personalities, exploring the interplay between individual traits, memory, and social interactions to replicate human-like processes, and trying to provide users with more engaging and natural experiences.

This paper introduces IMPECT-POSE, an innovative front-end and back-end architecture designed to enhance fitness and sports training through precise body posture tracking. This system integrates advanced computer vision and artificial intelligence in pose estimation to provide real-time feedback on exercise execution, which is crucial for maintaining proper technique, reducing injury risks, and optimizing training outcomes. Our evaluations, conducted at two distinct locations with multiple participants, demonstrate the system’s capability to improve exercise performance significantly. The system’s flexibility allows sports professionals to monitor and guide clients remotely, enhancing the accessibility and effectiveness of training regimens. This research highlights the potential of augmented intelligence in transforming sports training, offering a scalable and effective alternative to conventional methods, and paving the way for future advancements in AI-driven personalized training programs. The continued development of this technology aims to refine its accuracy, broaden its applicability to diverse user preferences, and extend its use in practical, real-world settings.

Research in the recommender systems field typically applies a rather traditional evaluation methodology when assessing the quality of recommendations. This methodology heavily relies on incorporating different forms of user feedback (e.g., clicks) representing the specific needs and interests of the users. While this methodology may offer various benefits, it may fail to comprehensively project the complexities of certain application domains, such as the news domain. This domain is distinct from other domains primarily due to the strong influence of editorial control in the news delivery process. Incorporation of this role can profoundly impact how the relevance of news articles is measured when recommended to the users. Despite its critical importance, there appears to be a research gap in investigating the dynamics between the roles of editorial control and personalization in the community of recommender systems. In this paper, we address this gap by conducting experiments where the relevance of recommendations is assessed from an editorial perspective. We received a real-world dataset from TV 2, one of the largest editor-managed commercial media houses in Norway, which includes editors’ feedback on how news articles are being related. In our experiment, we considered a scenario where algorithm-generated recommendations, using the K-Nearest Neighbor (KNN) model, employing various text embedding models to encode different sections of the news articles (e.g., title, lead title, body text, and full text), are compared against the editorial feedback. The results are promising, demonstrating the effectiveness of the recommendation in fulfilling the editorial prospects.

Environmental sustainability is the avoidance of the depletion of natural resources in order to maintain an ecological balance. To influence people to be environmentally sustainable, several mobile apps exist to teach sustainable behaviours. These apps hold great promise as interventions for influencing people to live sustainable lives since they are often designed using behaviour change strategies. However, the effectiveness of these apps as behaviour change tools is unclear. In addition, despite people downloading these apps, their engagement level is still low. Research suggests that user experience and usability determine the adoption and usage of apps. Research also indicates that user experience and usability of apps can be gleaned from the reviews written by users on the app store. Thus, reviews are a good source of determining the user experience and usability of users. To determine the effectiveness of sustainability apps as behaviour change tools and the users’ experience with the apps, we reviewed 70 sustainability apps that are available on the Google Play Store. First, using a popular behaviour change framework, the App Behaviour Change Scale (ABACUS), we investigated the persuasive strategies implemented by the apps and how these strategies were designed and implemented to achieve the targeted design objectives. Second, using natural language processing, we identified the common themes in the user reviews of the apps. The preliminary results presented here can influence the design of apps for influencing environmentally sustainable behaviour.

Many web-based systems such as online retail, information systems or search engines track the interactions users have with them. Tracked data can comprise high-level information like dwelling time, reviewed items, and clicked elements, but also fine-grained information in the form of mouse trajectories and keystrokes. While these data are often fed into user- or behavior models in recommender systems, there are few approaches for interactive visual exploration of multi-modal and complex interaction patterns. Yet, the thorough analysis could reveal important insights for the design and evaluation of said models. We propose a suitable visual analysis approach that allows to validate and correct models in an intuitive and interactive manner. Our tool provides insights into concrete user (inter)actions and also estimates more complex behavioral patterns. Level of detail views in our system outlines the certainty of detected behaviors and serve the explainability. Our approach can help engineers to understand user interactions and improve behavioral models.

We investigate the impact of user demographics (age, gender) and experience (with personas and chatbots) on users’ perceptions of interactive personas. A within-subjects study was conducted with 54 participants, mostly engineers and computer scientists. Each participant used interactive personas with two interfaces: a web-based profile persona and a chat persona. The findings from regression analysis indicate that users’ age and gender (as well as persona’s gender) affect multiple perceptions of personas. In addition, the interface modality (profile vs. chat) has a significant impact. Findings highlight the need for designing interactive personas that appeal to diverse user bases to increase the general accessibility of interactive personas. They also support the notion that the persona interface itself regulates user perceptions even when the persona’s information remains the same.

Piloting is a cognitively demanding task that not only requires conceptual knowledge and domain expertise, but also skills to intercept relevant visual cues in a changing environment. Pilots’ abilities to divide their visual attention effectively during critical flight maneuvers play a central role in their situation awareness and decision-making. Given that a large number of visual cues in the flight deck need to be processed by the human eyes, there is an opportunity to utilize eye tracking in the analytics of pilot performance. To this end, we present a case study involving 10 licensed pilots in a simulated flight takeoff scenario and demonstrate that specific gaze behaviors employed by a pilot may lead to varying levels of success. In particular, pilots who perform better during the climb phase may often exhibit faster visual searches, and those who perform better during the takeoff phase may often scan a larger area of the visual scene. Also, more successful pilots tend to report lower cognitive workload that is further supported by their pupil diameters in the experiment. This knowledge may be utilized in the development of user models to infer pilot success and failure, whereby an intelligent flight deck can adapt system behaviors to the pilot, whereby alerts, warnings, and automation may be initiated in the event of detecting irregularity in pilots’ gaze.

While repeating activities can create healthy habits, exploring new physical activities is also important to increase health benefits and prevent boredom. Following habit formation and variety-seeking behavior theories, this study investigates the difference between repetition and exploration of physical activities in health recommender systems. An eight-week Micro-Randomized Trial is conducted in which 11 physically inactive adults receive personalized activity recommendations that are either (a) a repetition or exploration, (b) connected to a specific location, duration, or neither, and (c) accompanied by videos or Points-of-Interest, or not. Analyses of the 187 submitted activity recommendations suggest that the inactive participants prefer exploration, as exploration recommendations were submitted the most, had significantly larger star ratings and durations, and are dependent on moderators (b) and (c). More specifically, exploration for workouts received the highest star ratings and motivation. To our knowledge, this study is the first to investigate repetition and exploration for physical activities, contributing to effective recommender algorithms for healthy behavior change.

Traditional knowledge modeling methods have primarily focused on student knowledge modeling using assessed learning activities, often overlooking the critical interplay between students’ knowledge and behavioral preferences. However, students typically interact with multiple types of learning materials, such as questions (assessed), video lectures (non-assessed), and textbooks (non-assessed). We argue that student knowledge can affect their behavioral preferences, and the choice of learning material type can influence their knowledge. In this paper, we address this gap by proposing a novel framework that models student knowledge and behavior as a multi-task learning problem with two objectives. Our dual objectives are to predict student performance and their preferences for selecting different types of learning materials. We utilize the Pareto Multi-Task Learning (MTL) algorithm to effectively handle the complexities of this multi-objective optimization, applying it to two advanced multi-activity knowledge modeling methods, TAMKOT and GMKT, which we refer to as Pareto-TAMKOT and Pareto-GMKT, respectively. We evaluate the framework on one real-world dataset. Our experimental results demonstrate that both Pareto-TAMKOT and Pareto-GMKT improve upon their original models and outperform all baseline models. This underscores the benefits of treating the modeling of student knowledge and behavior as a multi-task learning problem and addresses this multi-objective challenge through the application of Pareto MTL.

Lexical alignment is a form of personalization frequently found in human-human conversations. Recently, attempts have been made to incorporate it in human-computer conversations. We describe an experiment to investigate the trust of users in the performance of a conversational agent that lexically aligns or misaligns, in a collaborative task. The participants performed a travel planning task with the help of the agent, involving rescuing residents and minimizing the travel path on a fictional map. We found that trust in the conversational agent was not significantly affected by the alignment capability of the agent.

With increased technology use in education that leverages the benefits of user modeling, adaptation, and personalization, privacy of educational data gains relevance, yet privacy research in this area is still in its infancy. Current research on educational technologies focuses on the technological implementation and learning outcomes. Yet, potential privacy concerns or user perceptions that arise from the use of personal and sensitive data these systems require may impact user engagement, motivation, and learning. To explore potential predictors of students’ privacy-related risk perceptions, we conducted a study with N=66 university students who used an automated assessment software. The results show that awareness about the types of data the software collects and how it is used is associated with lower levels of risk perception and a higher preference for using the software. The findings present a first step of a research project aiming to provide a personalized learning experience while making related privacy implications graspable and controllable.

Human Robot Interaction (HRI) is a field of study dedicated to understanding, designing, and evaluating robotic systems for use by, or with, humans. In HRI there is a consensus about the design and implementation of robotic systems that should be able to adapt their behavior based on user actions and behavior. The robot should adapt to emotions, personalities, and it should also have a memory of past interactions with the user to become believable. This is of particular importance in the field of social robotics and social HRI. The aim of this Workshop is to bring together researchers and practitioners who are working on various aspects of social robotics and adaptive interaction. The expected result of the workshop is a multidisciplinary research agenda that will inform future research directions and hopefully, forge some research collaborations.

In this position paper we address the issue of acceptance in human-robot interaction, a topic often associated with anthropomorphism. The literature around this concept lacks of a discussion around the kinetic aspect of anthropomorphism, especially in applications where machine accuracy is required. We briefly summarize the results we achieved and outline some research directions promised to impact human-robot collaboration in sectors like manufacturing and medicine.

This paper presents an empirical study that examined how children interpret emotional body language displayed by the humanoid robot NAO. The purpose of the study is to provide insights into how children perceive and respond to emotional cues from robotic agents presenting an empirical evaluation that explores the effectiveness of using a humanoid robot to convey emotions to children. Through the examined results, the study aims to highlight the potential of using humanoid robots in educational and therapeutic contexts.

This project investigates the potential of personalized interactions with social robots, like Pepper, by customizing images and videos to align with individual user profiles. The primary goal is to make conversations with social robots more engaging and relatable. By tailoring content to factors such as age, gender, interests, and cultural background, we aim to enhance user satisfaction and foster broader acceptance of social robots in everyday settings. This study opens up new possibilities for designing robots that can adapt to diverse user needs and preferences. Ultimately, this approach could lead to social robots that are more widely accepted and valued in a range of contexts, from customer service to education to healthcare. This project provides a path for future research in context-aware robot communication.

Autonomous robots are expected to have safe and social interactions while sharing the environment and collaborating on several tasks with people. In such scenarios, robots’ behaviours must be transparent to people so that they can understand and predict the robots’ intentions and actions. In this work, we address the problem of transparency of a robot’s behaviours with particular attention given to the use of legibility of the robot’s actions, predictability of the final goal and the possibility of providing verbal explanations for a transparent human-robot interaction. Here, we also highlight different ways for adaptive generated explanations to be appropriately communicated to people and how to evaluate the overall transparency of the robots’ behaviours.

The following paper provides an overview of the “Feel Good” project, including its objectives, methodologies and plans. The premise of the project is to explore the integration of social robots within therapeutic frameworks aimed at easing the anxiety and distress associated with medical appointments for autistic youth. Social robots have shown effectiveness in supporting autism therapies, enhancing engagement and therapeutic outcomes. The aim of this initiative is to improve communication between healthcare providers and autistic youth. Through the implementation of various strategies, the initiative seeks to enhance the children’s capacity to identify and express their symptoms while also providing a comfortable healthcare experience. Furthermore, the initiative offers a secure environment where the children can work on their weaknesses and practice managing medical events. The activities are tailored according to their individual support needs and levels to ensure the best possible outcome.

The second edition of the Context Representation in User Modelling (CRUM) workshop, themed “Human-Centred Context”, aimed to provide a venue for research on contextual information within the evolving landscape of UMAP. This workshop explored context through a human-centred lens, emphasising the nuanced interplay between subjective and objective contexts in user experience rather than traditional domain-centred approaches. This theme encouraged accounting for the interaction between the user model and the contextual environment. The workshop also fostered discussion on the landscape of contextual representation – ranging from physical settings and social dynamics to technological interfaces – as well as on interaction design, user experiences, and influencing user behaviour. Prospective authors were invited to submit two types of papers: full papers of up to seven pages and two-page provocation papers. Accepted submissions are summarised in this workshop report.

Contextual information is a prerequisite for timely offering of personalized decision support and recommendation. Yet, research on context-aware recommender systems (CARS) does not appear to be thriving, and finding public datasets containing context factors is a challenging task. We can make various assumptions about why this drop in research interest happened – be it ethical considerations or the popularity of opaque deep learning models that merely consider context in an implicit way. This is an unwelcome development. We argue that continued effort must be put on the creation of suitable datasets. Furthermore, we see significant opportunities in the development of next-generation CARS in the space of interactive AI assistants powered by Large Language Models.

News media framing can shape public perception and potentially polarize views. Emotional language can exacerbate these framing effects, as a user’s emotional state can be an important contextual factor to use in news recommendation. Our research explores the relation between emotional framing techniques and the emotional states of readers, as well as readers’ perceived trust in specific news articles. Users (N = 200) had to read three economic news articles from the Washington Post. We used ChatGPT-4 to reframe news articles with specific emotional languages (Anger, Fear, Hope), compared to a neutral baseline reframed by a human journalist. Our results revealed that negative framing (Anger, Fear) elicited stronger negative emotional states among users than the neutral baseline, while Hope led to little changes overall. In contrast, perceived trust levels varied little across the different conditions. We discuss the implications of our findings and how emotional framing could affect societal polarization issues.

Adaptive and personalized systems have become pervasive technologies, gradually playing an increasingly important role in our daily lives. Indeed, we are now accustomed to interacting with algorithms that leverage the power of Language Models (LLMs) to assist us in various scenarios, from services suggesting music or movies to personal assistants proactively supporting us in complex decision-making tasks. As these technologies continue to shape our everyday experiences, ensuring that the internal mechanisms guiding these algorithms are transparent and comprehensible becomes imperative. The EU General Data Protection Regulation (GDPR) recognizes the users’ right to explanation when confronted with intelligent systems, highlighting the significance of this aspect. Regrettably, current research often prioritizes the maximization of personalization strategy effectiveness, such as recommendation accuracy, at the expense of model explainability. To address this concern, the workshop aims to provide a platform for in-depth discussions on challenges, problems, and innovative research approaches in the field. The workshop specifically focuses on investigating the role of transparency and explainability in recent methodologies for constructing user models and developing personalized and adaptive systems.

The growing exchange of opinions in online news forums brings together a diverse cross-section of users with varying opinions and motivations. Understanding these behaviors is crucial for unraveling the composition of these large user bases. This study proposes an explainable model aimed at classifying users based on their activity and interaction patterns in online news forums. The model leverages exploratory and statistical data analysis to reveal recurring behaviors and provides a tool to analyze the evolution of large user communities, offering an overview of their composition. The model identifies six active roles: Taciturn, Silent Voter, Regular, Conversationalist, Power User, and Celebrity, and one inactive role, Lurker. The model was evaluated for its predictive power, achieving a macro F1 score of 0.8632, demonstrating its robustness. By applying the model to a long-term dataset from the online news forum derStandard.at, an analysis of role distribution over time was conducted. The results indicated a gradual increase in user activity within the forum. Moreover, the study assessed the co-occurrence of roles in users’ long-term behavior and measured the frequency of role changes. This analysis aimed to determine whether users have consistent roles or exhibit various roles, which may depend on time or context.

Intelligent Personal Assistants (IPAs) have become ubiquitous through integration into smartphones, smart speakers, and standalone devices. However, prior studies raised noteworthy usability concerns and determined that IPAs remain primarily used for simple tasks. Such findings contrast with the reported user aspirations for more proactive and truly personalised IPAs. By focusing on the use case of time management, in this paper, we contemplate how scrutability – i.e. the ability of the user to study their assistant and its underlying user model – fits within the vision of more complex IPAs. Furthermore, we describe our ongoing project investigating user interest in and expectations of the scrutability of a proactive calendaring assistant. Lastly, by deliberating on the challenges and benefits of making IPAs scrutable, this paper outlines potential avenues for further research.

Open user models provide affordance for a transparent user control over recommendations based on shared symbolic representation within the system. Users must build their user profile by adding these symbols and tuning their importance to get meaningful recommendations. Since the link between these symbols and the reference explanation is often unavailable, it can be difficult for users to understand them. These symbols are often referred to as concepts, tags, areas, topics, labels, features, or keyphrases. This study showcases an information exploration system that helps students identify potential faculty members to collaborate with. The system works by matching user and faculty profiles that contain keywords or phrases representing topics/areas of interest. Students must develop their understanding of research topics while building their profiles, which can become challenging as they add more keywords. To support students in controlling the recommendation, we introduce post hoc explanations with three levels of detail: no explanations, individual explanation for topics, and explanation of the relationships between topics. This study explores how explanation is associated with the user context / tasks and the exploration process. Our observation suggests that expertise in the field is linked to exploring fewer novel topics and seeking fewer explanations but engaging more with explanations of relationships. In addition, we found that the engagement with faculty information is moderately correlated with the use of more advanced explanations.

In working environments where prolonged sitting is ubiquitous, maintaining correct posture is crucial to alleviating musculoskeletal problems and improving general well-being. This research presents an innovative approach for assessing sitting postures using any camera conveniently accessible to the subject, be it integrated into their personal computer or an external device. The main contribution of our system is to detect key posture points and incorrect postures and provide the final user with personalized feedback and explanations to help them correct their postural alignment. Using a simple architecture i.e. the multilayer perceptron, we succeeded in identifying human postures. Moreover, by evaluating the results obtained from our classification model, we are able to obtain an explanation based on a post-hoc analysis. We observed that the most impactful joints for correct posture are, in order of importance, the wrists, elbows, and shoulders. Inspired by counterfactual explanations, we consequently provide personalized feedback for users with incorrect postures. These outcomes, although preliminary, show that the proposed pipeline is encouraging and can be pursued in future work where increasing the variety of data and improving detection approaches will be predominant. The reproducibility code is available here: https://github.com/GaetanoDibenedetto/Explainable-Corrective-Feedbacks

Users are often confronted with situations where they have to decide in favor or against an offered item, like a book, movie, or recipe. Those suggested items are commonly determined by a recommender system, which considers personal preferences to identify relevant items. However, those systems often lack transparency and comprehensibility in revealing why a specific item is recommended. For this purpose, explanations have been added as a powerful tool to help users with their final decisions. In this paper, we present and evaluate the capabilities of a Large Language Model (LLM) to come up with high-quality explanations to further improve the support of users for three different recommendation approaches, including feature-based recommendation, collaborative filtering, and knowledge-based recommendation. We explain how an LLM can be applied to generate personalized explanations and evaluate the explanation goals in an online user study. Our findings highlight that LLM-generated explanations are highly appreciated by users as they help in the evaluation of recommended items. Furthermore, we discuss which characteristics of the LLM-based explanations were perceived positively and how those findings can be used for future research.

News Recommender Systems (NRSs) have become increasingly pivotal in shaping the news landscape, particularly in how news is disseminated. This has also led to concerns about information diversity, especially regarding selective exposure in the realm of political news. Users may not recognize that news content presented to them is subject to selective exposure, through users that incorporate political beliefs. Within the U.S. two-party system, our research explores the interactions between NRSs and users’ ability to discern news articles that align with their political biases. We performed an online experiment (N = 160) to address the issue of user awareness and self-recognition of selective exposure within NRSs. Users were asked to select any number of news articles that matched their political orientation (i.e., Democrat or Republican) from a list of 50 news articles (5 Democrat, 5 Republican, 40 filler articles), which were either ranked saliently towards their political orientation or randomly. Contrary to expectations, our findings reveal no significant difference in article selection between participants exposed to a baseline random order and those who where presented with the more salient and easy to select version. We did observe that Republicans performed worse than Democrats in identifying aligning articles, based on precision and recall metrics.

Recommender systems can help web users find more relevant content, improve their online experience, and support them in the discovery of new Points-of-Interest (POI). Yet, challenges persist in dealing with the cold-start problem and in recommendation explainability. To address these, we have created ScrollyPOI, an interactive POI recommender system based on Data Humanism principles. Utilizing scrollytelling, we address the cold-start problem by engaging users in reflecting on previous positive experiences. Additionally, ScrollyPOI enhances explainability through input and output explanations. The system uses stacked bar charts and word clouds to explain how user preferences inform recommendations (input). Finally, ScrollyPOI employs a multi-layered approach to explain why specific POIs are recommended (output). We have evaluated ScrollyPOI’s interface and experience through a preliminary study, highlighting its potential for transparent explanations in the POI recommendation domain. Our findings underscore ScrollyPOI’s efficacy in collecting preferences and enhancing recommendation transparency, positioning it as a platform for studying explainability goals in the POI domain.

Conversational recommender systems (CRS) are increasingly able to conduct complex conversations with the user due to the advances in natural language processing techniques. As a consequence, dialogs in CRS can go beyond the mere recognition of product-related intents, offering decision aids that let users, for example, compare or critique recommendations. Users may, however, not be aware of such functionality and fail to actively query the system for such functions. It may thus be helpful, if the system proactively provides cues for using the advanced functions. Whether and how such proactive cues are perceived and used by the user has not been investigated yet. We report an online study investigating user interaction behavior under two interaction schemes: a proactive scheme prompting the user to use the functions and a passive scheme without prompts where users can freely enter their requests. To compare the two schemes, we implemented a chatbot in the domain of smartphones. Two groups of participants on Prolific (total n=270) used the system, which operates either with a proactive or passive scheme. In addition to interaction data, we measured several psychological factors and users’ subjective assessment of the system through questionnaires. We mainly found: 1. There is no significant difference in user experience under the two schemes. 2. Users tend to accept prompts when the system provides them. 3. Overall, more intuitive users (as measured by an established decision-making style instrument) tend to accept the system’s prompts more than more rational persons. 4. A tendency that acceptance of interaction prompts aligns with self-reported dialog-related preferences. The results provide heuristic suggestions for dialog strategy design and the potential personalization of CRS.

While most recommender systems cater to individual users’ needs, there are numerous situations where these systems are needed to meet groups’ demands. These systems are broadly labelled as Group Recommender Systems (GRSys). Traits like interpersonal relationships, group mood, and emotional contagion are essential to fulfilling the group’s needs. However, the group’s characteristics are frequently ill-defined and dynamic and are typically absent from systems modeling. Moreover, GRSys must maneuver between the needs of the group and the individuals when opinions differ and can contradict each other. The third edition of GMAP proposes consolidating a community of scholars interested in group modeling, adaptation, and personalization. Through the workshop, researchers continue their examination of the difficulties and possibilities of creating efficient procedures and instruments to facilitate collective decision-making. GMAP 2024 offered this unique opportunity to gather scholars from different fields to enrich discussions over GRSys’ research. The workshop also allowed attendees to strengthen their networks and establish new connections conducive to cutting-edge collaborative research.

When groups of people conduct recommended activities together, several factors affect their individual satisfaction. This includes the satisfaction of the other group members, due to phenomena such as interpersonal social influence and emotional contagion. To model the change in individual satisfaction in a group context, it is important to consider factors like the individual personality and the type of relationships between group members. This research aims to investigate how individual satisfaction related to a group activity is influenced by the preferences of other people present. We focus on groups with two members, in the context of listening to music. We performed an exploratory user study (n=26; 13 pairs), finding that people adjust their preferences to their companion, but we did not see an effect of the impact of tie strength (friend vs. stranger). We also found indications that personality traits can predict whether participants are more (Dominating and Avoiding), or less likely (Agreeableness) to adjust their preferences to their companion. Although influenced by the limited sample size, our results support the hypothesis that individual personalities impact the individual satisfaction shift. However, measuring different phenomena that can affect users’ behaviors in the considered experimental setting seems crucial, and deeper investigations with a larger sample size are necessary to produce more robust results.

In a decision-making scenario, one group member may need, independently from the other members, to choose an item, e.g., a restaurant, that will be experienced by the group. In a prior research on restaurant recommendation, we have identified two primary tasks of such an organizer of a lunch/dinner event, who is in charge of selecting a proper restaurant for the group: anticipating the other group members’ preferences, to properly enter these preferences in the recommender system, and reconciling incompatible preferences, if they arise in the elicitation phase. To support the first task, we augmented a group recommender system with a machine learning model that predicts group members’ food preferences, about dishes that can be consumed in the restaurant, based on other available information about the members (demographics and preferred cuisine). However, in a user study, we found that supporting functionality to be not effective in improving the quality of the choices made by the organizer.

In this paper, we investigate the causes of this poor performance. We analyze the possibility that the poor performance may relate to: the deployed ML models used for food preference prediction, the amount of data about food preferences used for training the ML model, and the complexity of the preference anticipation task. The results of our experiments suggest that neither the ML model nor the scarcity of the preference data is responsible for the observed poor value of the implemented preference recalling function. While a user study seems to confirm the impossibility of accurately predicting food preferences from the considered user’s characteristics (demographics and preferred cuisine).

Beyond accuracy metrics, such as fairness and diversity, have become widely studied topics in recommender systems. Improving these metrics is important not only from an ethical and legal perspective, but can also improve overall user satisfaction. Although these metrics are widely discussed, very little empirical research has been done, especially comparing multiple algorithms across different metrics. This work explores the role of fairness and diversity in news recommender systems, specifically in the context of the Austrian media landscape. This study aims to identify the most effective approaches for generating fair and diverse news recommendations, while addressing the potential negative consequences of biased recommendations and filter bubbles, such as societal polarization and the suppression of information. This includes an extensive literature review of relevant group unfairness metrics and state-of-the-art fairness-aware algorithms. A dataset of articles from an Austrian newspaper was used for empirical research, with analysis performed on fairness, and diversity of recommendations. The key message of the study is that accuracy and fairness can be achieved simultaneously with the right modeling approach, while diversity can be held constant using these modeling techniques. The study recommends the use of Personalized Fairness based on Causal Notion models for accuracy and reducing certain unfairness metrics, and finds Fairness Objectives for Collaborative Filtering models more effective at reducing other types of unfairness. The findings contribute to the field by demonstrating the importance of incorporating these metrics into the design and evaluation of recommender systems.

Teamwork is prevalent in many settings including in higher education. Unfortunately, teamwork does not always go well. We previously studied how a peer assessment survey can be used by individuals to give their opinions on their teammates. This allows somebody (such as a teacher in an educational setting) to monitor how well the teams are doing. We also studied how to provide support to the person who fills out the survey and adapt such support to their personality and the ratings they have given. This paper focuses on another aspect, namely what feedback to provide to a person who has been rated by their teammates. We discuss the challenges of providing such feedback automatically and provide a research agenda for solving this issue.

User experience, which encompasses users’ feelings and perceptions, is regarded as a key element in the evaluation of recommender systems. The existing literature extensively works on recommendation generation strategies with focus on the accuracy by considering objective aspects of the system. Although some of the current works considered subjective aspects of the recommendation systems from a user-centric perspective to evaluate the recommender system, however, a comprehensive analysis that could investigate factors to improve user experience was of limited focus. In this paper, we propose a methodology that provides a comprehensive multi-perspective analysis of a social-aware fashion recommender system and analyses the impact of user’s personal attributes and profiles on their experiences in various aspects of system use. A user study was conducted to realize the proposed methodology. The obtained insights highlighted that user experiences vary not only from the perspective of using a recommender system but also by varying their personal attributes (age, gender, hobby) and profiles.

Nowadays, the profound digital transformation has upgraded the role of the computational system into an intelligent multidimensional communication medium that creates new opportunities, competencies, models and processes. The need for human-centered adaptation and personalization is even more recognizable since it can offer hybrid solutions that could adequately support the rising multi-purpose goals, needs, requirements, activities and interactions of users. The HAAPIE workshop1 embraces the essence of the “human-machine co-existence” and brings together researchers and practitioners from different disciplines to present and discuss a wide spectrum of related challenges, approaches and solutions. In this respect, the ninth edition of HAAPIE includes 4 long papers and 2 short papers.

Students in higher education with Autism Spectrum Disorder (ASD) face many challenges that might differ from their neurotypical peers. One area where students with autism face deficits is planning. This paper looks into the challenges faced and the strategies applied for planning by students with autism compared to neurotypical students. We aim to identify where personalisation and adaptivity may help them become independent and effective in their planning behaviour. This research indicates which personalisation needs designers of assistive technologies should consider for planning and task management. We present an online survey with 30 neurotypical students (NTS) and 34 students with (self-)diagnosed autism (ASDS) and interviews with six students with autism for more in-depth insights. Results indicate that ASDS experience problems with gaining a clear overview of what they need to do, knowing when they might do it, and following that plan to execution. In contrast to NTS, they also struggle with fitting routine household and self-care tasks into their schedule. We identified time-independent planning, identifying external pressure and defining sub-tasks as promising planning strategies for students with autism when combined with adaptation to their personal needs.

Students experience motivational issues during online learning which has led to explorations of how to better support their self-regulated learning. One way to support students uses social reference frames or social comparison in student-facing learning analytics dashboards (LADs) and open learner models (OLMs). Usually, the social reference frame communicates class averages. Despite the positive effects of class-average-based social comparison on students’ activity levels and learning behaviors, comparison to class average can be misleading for some students and offer an irrelevant reference frame, motivating only low or high performers. Such conflicting findings highlight a need for an investigation of social reference frames that are not based on the “average” student. We extend the research on social comparison in education by conducting two complementary classroom studies. The first explores the effects of different fixed social reference frames in a non-mandatory practice system, while the second introduces an adaptive social reference frame that dynamically selects the peers who serve as a comparison group when students are engaged in online programming practice. We reported our analyses from both studies and shared students’ subjective evaluations of the system and its adaptive comparison functionality.

As the demand for personalized and adaptive learning experiences increase, there is a urgent need for providing effective feedback mechanisms within critical systems, such as in psychomotor learning systems. This proposal introduces an approach for the integration of retrieval-augmented generation tools to provide comprehensive and insightful feedback to users. By combining the strengths of retrieval-based techniques and generative models, these tools offer the potential to enhance learning outcomes by delivering tailored feedback that is both informative and engaging. The proposal also emphasises the importance of incorporating explainability and transparency concepts. Following the hybrid intelligence paradigm it is possible to ensure that the feedback provided by these tools is not only accurate but also understandable to humans. This approach fosters trust and promotes a deeper understanding of the psychomotor learning process, empowering users and facilitators to make informed decisions about the psychomotor learning path. The hybrid intelligence paradigm, which combines the strengths of both human and artificial intelligence, plays a crucial role in the deployment of these solutions. By taking advantage of the cognitive capabilities of human experts alongside the computational power of artificial intelligence algorithms, it is possible to offer personalised feedback that takes into account both technical accuracy and pedagogical effectiveness. Through these collaborative efforts it is also possible to create learning environments that are inclusive, adaptable, and beneficial to lifelong learning. In conclusion, this proposal introduces retrieval-augmented generation tools for providing feedback in psychomotor learning systems, which represents a significant step towards in its personalization, and whose ethical implications align with the new regulations on the implementation of intelligent technologies in critical systems.

Recommender systems play a crucial role in our daily lives, constantly evolving to meet the diverse needs of users. As the pursuit of improved user experiences continues, metrics such as serendipity have emerged within the realm of beyond-accuracy paradigms. However, integrating serendipitous recommendations presents complex challenges, necessitating a delicate balance between novelty, relevance, and user engagement. In this interdisciplinary experimental user study, we address these challenges within the context of a book recommender system. By investigating the impact of interface design changes on user trust, a key determinant of satisfaction with serendipitous recommendations, we measured trust levels for both individual recommended items and the recommender system as a whole. Our findings indicate that while interface enhancements did not yield significant increases in trust, they did notably elevate serendipity ratings for previously unknown books. These results highlight the intricate interplay between technical and psychological factors in the design of recommender systems, emphasizing the importance of human-centered approaches in the creation of more responsible AI applications. This research contributes to ongoing discussions surrounding user-centric recommendation systems and aligns with broader themes of digital humanism and responsible AI.

This research explores the integration of Human-in-the-loop (HITL) control within Proactive Intelligent Personalised Agents (PIPAs) that possess the capability to proactively anticipate users’ needs and perform tasks on their behalf. The proactive assistance offered by PIPAs is tailored to individual users’ preferences and behaviours. However, it is crucial to personalise the level of proactivity exhibited by PIPAs to align with users’ preferences regarding the potential delegation of autonomy. This necessitates HITL control to regulate PIPAs’ autonomy levels, ensuring appropriate user involvement in decision-making. Through a simulation-based approach, this research explores the integration of HITL control in PIPAs. It investigates the conditions that trigger HITL control, its mechanisms, and the challenges associated with these triggers.

The advent of modern information technology such as Large Language Models (LLMs) allows for massively simplifying and streamlining the communication processes in human-machine interfaces. In the specific domain of healthcare, and for patient practice interaction in particular, user acceptance of automated voice assistants remains a challenge to be solved. We explore approaches to increase user satisfaction by language model based adaptation of user-directed utterances. The presented study considers parameters such as gender, age group, and sentiment for adaptation purposes. Different LLMs and open-source models are evaluated for their effectiveness in this task. The models are compared, and their performance is assessed based on speed, cost, and the quality of the generated text, with the goal of selecting an ideal model for utterance adaptation. We find that carefully designed prompts and a well-chosen set of evaluation metrics, which balance the relevancy and adequacy of adapted utterances, are crucial for optimizing user satisfaction in conversational artificial intelligence systems successfully. Importantly, our research demonstrates that the GPT-3.5-turbo model currently provides the most balanced performance in terms of adaptation relevancy and adequacy, underscoring its suitability for scenarios that demand high adherence to the information in the original utterances, as required in our case.

While digital transformation brings broad positive impacts, its rapid propagation varies across territories, accentuating disparities not only in the social and organisational contexts but even in industrial sectors. Acknowledging this digital divide, the HUMAD workshop aimed to reshape industrial and digital landscapes for equity and accessibility, specifically emphasising the pivotal role of user modelling and artificial intelligence. Discussions centred on personalised user modelling and holistic, human-centred approaches tailored for an industrial context. The workshop explored applications in critical sectors, including smart cities, tourism, industrial production, healthcare, education, and well-being, focusing on delivering quality digital interactions and services tuned to industrial and human-centric needs. The objective was to elevate the quality of digital interactions and services, ensuring a more uniform and inclusive distribution of the benefits of industrial digital transformation. Beyond immediate concerns, HUMAD delved into user modelling’s potential to drive growth in disadvantaged territories. By comprehending the unique challenges these regions face, the aim was to cultivate sustainable solutions within an industrial framework. In the light of the above, we invited contributions that navigate the intricacies of user modelling, proposing innovative strategies to reduce disparities and develop a more equitable, human-centric digital future.

Smart sensory devices, such as smart watches, scales, and blood pressure gauges, are increasingly adopted by individuals aiming to improve their health and fitness. Those devices gather extensive data about cardiovascular parameters, physical activities, sleep quality, and behavior. Thanks to data analytics and artificial intelligence algorithms, they provide insights into the health status of individuals. Derived data is used to support self-care interventions and to provide practitioners with additional health information acquired on a continuous basis. However, most of the current solutions focus on the physical dimension of health, while the mental dimension is often neglected. In this paper, we present the initial investigation of a system to recognize a wide range of psychological parameters, including behavioral inhibition/activation, anxiety, and stress, leveraging data acquired from personal healthcare devices. We experimented with the application of different supervised learning algorithms on features extracted from heart, sleep, and inertial sensor data acquired from a cohort of 21 individuals over 24 hours each. Our preliminary findings suggest that our method may yield promising outcomes in recognizing different aspects of mental well-being. However, due to the limited size of the used dataset, a more comprehensive experimental evaluation, with a broader number of participants and carried out over an extended monitoring period, is imperative to substantiate the results.

The need for recycling and a circular economy has grown significantly in recent years, accompanied by challenges such as greenwashing and fraudulent recycling practices. Blockchain technology is a new emerging technology that aims to facilitate information sharing, enable comprehensive government oversight, and establish effective incentive mechanisms. This paper explores digital transformation within blockchain frameworks to revolutionize waste management certification processes. Through the description of a real-world scenario, involving the recycling of food containers, in our case study pizza boxes within a pizzeria, this study highlights the tangible benefits, and challenges encountered from integrating blockchain technology into waste management processes. This case study demonstrates the potential for technology-driven solutions to address environmental challenges while fostering community engagement and incentivizing sustainable practices with the digital transformation of SMEs.

In this paper, we present our currently on-going work on a method for analyzing digital health transformation in our society by constructing a Knowledge Graph from a large corpus of 7.8 million English news articles, dating from 1987 through 2023. We firstly sampled around 95k articles relevant to the Digital Health topic by training and deploying a Deep Learning binary classifier via fine-tuning BERT. Successively, by deploying NLP techniques, we extracted triples from the identified articles to form a Digital Health News Knowledge Graph, which consists of 431k distinct triples connecting 186k entities through 1866 relations. The constructed Knowledge Graph provides insights into the evolution of Digital Health in news media and serves as a resource for further research in the field. The analysis that we have carried out reveals significant trends in Digital Health as reflected in the news, with notable peaks coinciding with key events like the COVID-19 pandemic.

In recent years, the rapid evolution of smart technologies has spurred enterprises to undergo digital transformations, revolutionizing their business processes and operations. This shift, known as Digital Transformation, has permeated diverse sectors, particularly impacting production systems. Notably, Artificial Intelligence (AI) and robotic automation have emerged as pivotal drivers in this transformation, promising enhanced efficiency and innovation in industrial digitization. This paper presents a novel architecture designed to facilitate digital transformation within enterprises, harnessing the capabilities of advanced collaborative robots (cobots) and cutting-edge image segmentation techniques. Focused on a practical scenario within a food production environment, our proposed architecture aims to seamlessly integrate a cobot and a camera in an automatic system for efficient cardboard disposal. Specifically, our attention is drawn to the challenge of differentiating sections of food packaging suitable for disposal from those contaminated with stains or organic residues, a task with significant implications for waste management efficiency. By leveraging a cloud-based architecture and deploying AI algorithms for image segmentation, localization, and robot guidance, our study showcases the tangible benefits and practical applicability of these methodologies in real-world settings. This research not only highlights the potential of AI-driven solutions in addressing specific industrial challenges but also underscores the broader impact of digital transformation on optimizing operational processes and driving innovation across sectors.

Digital innovation is an important and evolving process that refers to the transformation and integration of digital technologies into various aspects of everyday life and work. In this context, innovative digital solutions are often adopted and developed, including the Internet of Things (IoT). In order to fully understand the potential, trends, and challenges related to the implementation of these technologies in data management, we seek to identify the most promising architectural choices for an IoT infrastructure from the perspective of key aspects in the context of Data Management. In this way, we aim to provide a comprehensive overview of IoT systems to guide organizations towards more informed decisions, offering cutting-edge solutions and contributing to digital transformation. This work aims to examine aspects such as Data Collection, Data Aggregation, Data Integration, Data Security, Data Retention, and Data Analysis, with particular attention to emerging challenges in these contexts. Within the context of digital innovation, our study focuses on several areas of interest. For example, in the field of smart cities, we explore how the use of Cloud, Fog, and Edge Computing technologies can improve video surveillance and promote the use of Artificial Intelligence (AI) for better management of smart cities. Cloud computing, characterized by centralized data processing and storage, represents a well established paradigm of IoT infrastructure. Fog computing extends these capabilities to the network edge, bringing computation and data storage closer to the data source, thereby reducing latency and enhancing real-time processing. Edge computing takes this concept further by processing data directly at or near the data source, minimizing the need for data to traverse long distances to centralized hubs, thus enabling even faster response times and improved efficiency. Additionally, we examine how human-centric methods and user modeling tools can facilitate the transition to intelligent environments, such as workplaces, healthcare, and cities. Considering the emerging challenges in IoT data management, our work provides an overview of the solutions offered by Cloud, Fog, and Edge Computing, enabling industries and small and medium-sized enterprises to adopt targeted and sustainable digital transformation strategies.

In the era of digital transformation, the utilisation of Virtual Reality (VR) applications in brain injury rehabilitation presents a novel frontier with vast potential. However, concerns persist about how traditional communication pathways between therapists and patients can be transformed into digital equivalents. Immersive virtual reality submerses patients in a virtual world, with therapists taking the role of external spectators. This has implications for the therapeutic alliance (TA), which is based on interpersonal bonds and effective communication between therapist and patient. This study investigated the use of immersive VR in clinical settings for the rehabilitation of acquired brain injury (ABI). The aim was to understand how to design a smart virtual therapeutic system that can be adapted to various needs of patients and therapists during the intervention. Through observation and semi-structured interviews, patient-therapist interactions were analysed during VR-assisted and traditional rehabilitation sessions, revealing challenges related to effective collaboration. Findings indicated changes in interaction and communication with the introduction of VR: informal dialogue decreased, while shared laughter and therapist feedback increased. Despite the visual disconnect imposed by VR headsets, therapists continued to use guiding gestures and engaged with patients. Interviews highlighted the importance therapists place on rapport-building, and both relied on each other for expertise and reassurance. This research offers valuable insights into the potential to transit, transform and enhance traditional rehabilitation practices with the use of virtual reality and suggests avenues for human-centred design of VR applications tailored to ABI rehabilitation and TA, ultimately enhancing therapeutic outcomes in this domain.

Behaviour change, crucial in health interventions, demands an understanding of the end user’s psychological, social, and environmental contexts, often overlooked in user modelling. This paper advocates integrating behaviour change theories into user models to improve personalisation based on individual characteristics. We introduce a novel, enriched with psychological theories, framework for using dynamic user Personas within Digital Behaviour Change Interventions, allowing adaptation to evolving user behaviours and preferences. The dynamic user Personas offer a detailed representation that accounts for the complex stages of behavioural change and COM-B patterns. We also explore the transition in user modelling from expert-driven to data-driven methods, highlighting the importance of acknowledging both interpersonal and intrapersonal variations in user behaviours. Our proposed semi-automated system model merges these modelling methods, enhancing personalisation in health interventions by addressing human behavioural complexities. Additionally, we discuss the ethical and privacy issues involved, ensuring responsible and secure data management. In conclusion, our approach conceptually links behaviour change, interaction design, and user modelling, setting a foundation for further empirical research and digital health applications.

Utilizing monocular cameras for 3D object understanding is widely recognized as a cost-effective approach, spanning applications such as autonomous driving, augmented/virtual reality or roadside monitoring. Despite recent progress, persistent challenges arise in creating generalized models adaptable to unforeseen scenarios and diverse camera configurations. In this work, we focus on the task of monocular 3D object detection within roadside environments. To begin, we introduce a versatile methodology for generating and labeling datasets tailored to roadside scenarios, addressing limitations encountered in real-world settings. Subsequently, we develop an array of deep learning models tailored to this task, refining them to address practical challenges that emerge during real-world application. Lastly, leveraging our framework, we curated a synthetic benchmark dataset comprising 1,415,680 frames and 8,902,636 labeled 3D objects, ultimately assessing the performance of existing models across all datasets.

In the current digital transformation scenario, Knowledge Graphs (KGs) represent an across-the-board instrument for representing knowledge in a structured form. Such tools allow to effectively enhance the performance of Artificial Intelligence models in manifold contexts, such as reasoning or information retrieval. Nevertheless, the effectiveness of KGs is often affected by the incorrect directionality of some of their edges, due in most cases to human error or the inefficiency of automatic and semi-automatic graph creation methods. This paper proposes a classification-based approach to identify misdirected triples within a KG, aiming to support and assist humans in creating graph refinement. Triples are the main component of KGs, and they model the connection between nodes with a <subject, predicate, object> form. Our proposal allows us to refine a KG by devising a classification-based approach for recognizing whether the subjects and objects are not compliant with the logic directionality of the corresponding predicate, meaning that they should be switched (e.g., the triple <U.S.A., is capital, Washington> should be inverted as <Washington, is capital, U.S.A.>). We compare traditional machine learning techniques with cutting-edge advanced methods, including pre-trained language models and large language models. Extensive experiments have been performed across several datasets, confirming the effectiveness of our proposal.

In the current Digital Transformation scenario, Knowledge Graphs are essential for comprehending, representing, and exploiting complex information in a structured form. The main paradigm in automatically generating proper Knowledge Graphs relies on predefined schemas or ontologies. Such schemas are typically manually constructed, requiring an intensive human effort, and are often sensitive to information loss due to negligence, incomplete analysis, or human subjectivity or inclination. Limiting human bias and the resulting information loss in creating proper Knowledge Graphs is paramount, particularly for user modeling in various sectors, such as education or healthcare. To this end, we propose a novel approach to automatically generating a proper entity schema. The devised methodology combines the language understanding capabilities of LLM with classical machine learning methods such as clustering to properly build an entity schema from a set of documents. This solution eliminates the need for human intervention and fosters a more efficient and comprehensive knowledge representation. The assessment of our proposal concerns adopting a state-of-the-art entity extraction model (UniNER) to estimate the relevance of the extracted entities based on the generated schema. Results confirm the potential of our approach, as we observed a negligible difference between the topic similarity score obtained with the ground truth and with the automatically generated schema (less than 1% on average on three different datasets). Such an outcome confirms that the proposed approach may be valuable in automatically creating an entity schema from a set of documents.

Following the successful series of PATCH workshops, PATCH 2024 will be again the meeting point between state-of-the-art cultural heritage research and personalization – using any technology, while focusing on ubiquitous and adaptive scenarios, to enhance the personal experience in Natural and Cultural Heritage sites. The workshop aims to bring together researchers and practitioners interested in exploring the potential of ICT technology (onsite and online) to enhance the visit experience. The expected result of the workshop is sharing and discussing novel ideas and creating a multidisciplinary research agenda that will inform future research directions and hopefully, forge some research collaborations. This summary provides an overview of the papers accepted for presentation at the workshop and publication in its proceedings.

Eye-tracking technology has gained prominence in cultural heritage studies, facilitating behavioral analysis and visitor engagement assessments. This paper explores the challenges and future directions of artwork segmentation in eye-tracking experiments, aiming to automate the identification of areas of interest. Although existing segmentation approaches, such as semantic segmentation models, show promise, they face limitations in accurately segmenting diverse artwork styles. We propose hybrid segmentation as a viable strategy, combining multiple techniques for improved accuracy. Through qualitative analysis, we evaluate segmentation models on public domain artworks, highlighting the strengths and weaknesses of each approach.

The cultural heritage of a place is mainly presented in websites and guides that describe Points of Interest focusing on specific topics or temporal windows. As such, they offer a detailed but fragmented viewpoint of the territory they describe and weakly support the provision of a longitudinal perspective on places and people’s lifestyles.

In the Revisualizing Italian Silentscapes (RevIS) project, we aim to achieve this perspective by exploiting cinematographic production through the multifaceted exploration of films and their scenes in a movie catalog. Providing advanced search support enhances the preservation and fruition of the cultural heritage related to early and recent movies. However, in this paper, we focus on exploring Points of Interest through the description of landscapes and related places or locations in films, a novel aspect we introduce. Specifically, we present a module of the RevIS app supporting the location-based exploration of movies.

Tourism Recommender Systems (TRS) have evolved from only providing user recommendations to becoming convergence points for multiple stakeholders. This necessitates recognizing the interests of all stakeholders, particularly in the tourism sector, which faces challenges like seasonality and resource constraints. Our stakeholder classification identifies consumers, item providers, platform, and society as key stakeholders, highlighting the complexity of real-world relationships. Fairness in TRS demands a multistakeholder approach, integrating sustainability to address the broader societal impact.

While research has focused on fair recommendation systems in tourism, the focus on generating sustainable recommendations remains limited. This demo paper aims to enhance fairness in TRS, mainly focusing on society as a stakeholder. We introduce the Green Destination Recommender (GDR), an application that prioritizes Societal Fairness (S-Fairness) by encouraging environmentally conscious decisions. GDR recommends sustainable tourism destinations based on the user’s starting location, travel month, and specific interests. The application promotes ecologically friendly options by recommending less popular yet appealing destinations, considering the emissions from transport to reach the destination and the seasonal demand to balance visitor numbers year-round.

With the increasing attention to environmental sustainability, matching user preferences and green behavior has become a must in several domains, including tourism. However, changing people’s traveling habits is hard and requires a relevant motivation effort. This short paper investigates the exploitation of digital nudges to promote sustainable tourism in personalized mobile guides for natural and cultural heritage exploration. The ultimate goal is to enhance point-of-interest recommender systems with the capability to drive users toward the selection of itineraries that they like and that can be managed by exploiting green means of transportation. For this purpose, we propose to integrate the recommendation of Points of Interest that satisfy the user’s interests with an explicit presentation of the environmental impact of traveling to such places, using digital nudges to drive the user toward the selection of sustainable tour management solutions.

Galleries, libraries, archives, and museums (GLAM) are distinct but interconnected institutions that play crucial roles in preserving, studying, and sharing knowledge, culture, and heritage (GLAM). Many hold and preserve unique artifacts that they make available to the public in various forms, on-site through exhibitions and displays, and remotely, through digital surrogates such as images, audio/video files, in digital exhibits and through various forms of description such as catalog, archival, and museum records management systems. Oftentimes these artifacts and resources need be delivered to the public in varying modes and for varying reasons: no one description suits all occasions and purpose. Short and accurate descriptions of. GLAM resources and artifacts are vital to the work of public engagement. The appearance of large language models and interfaces to interact with them such as Chatgpt4 opens new opportunities for automatic content creation, while posing also new challenges. In this position paper, we propose to use these new tools as an aid for the curator to create suggestions for content that may be used as descriptions of artifacts while enabling also the adaptation of the content for varied audiences and even to personal preferences.

Understanding the visitors’ behaviors can shape more immersive museum visits by providing personalized experiences. The first step to providing personalized experiences is to identify visitor types with common visit behavior characteristics. Several attempts have been made with promising results focusing on physical museums. Given the rise of virtual museum visits, in this paper, we take the first step toward identifying visitor types in virtual museums. To this end, we leverage data captured while visitors move within a virtual museum and interact with its exhibits. Using a machine learning approach, we identified four well-defined and distinct clusters that describe different types of virtual museum visitors.

The proliferation of AI-based techniques poses a range of new challenges for the design and engineering of intelligent and adaptive systems since they tend to either act as black boxes or, more generally, not offer the user sufficient transparency, control, and interaction opportunities, which are considered major goals of user-centred design in the HCI field. This workshop aims to share and discuss recent developments at the intersection of HCI and AI and to explore novel methodological, technical, and interaction approaches. Researchers and practitioners with diverse disciplinary backgrounds can and should contribute to addressing the challenges in this emerging field of human-centred artificial intelligence.

Artificial Intelligence (AI) systems involve diverse components, such as data, models, users and predicted outcomes. To elucidate these different aspects of AI systems, multifaceted explanations that combine diverse explainable AI (XAI) methods are beneficial. However, popularly adopted user-centric XAI evaluation methods do not measure these explanations across the different components of the system. In this position paper, we advocate for an approach tailored to evaluate XAI methods considering the diverse dimensions of explainability within AI systems using a normalised scale. We argue that the prevalent user-centric evaluation methods fall short of facilitating meaningful comparisons across different types of XAI methodologies. Moreover, we discuss the potential advantages of adopting a standardised approach, which would enable comprehensive evaluations of explainability across systems. By considering various dimensions of explainability, such as data, model, predictions, and target users, a standardised evaluation approach promises to facilitate both inter-system and intra-system comparisons for user-centric AI systems.

Biases in Artificial Intelligence (AI) or Machine Learning (ML) systems due to skewed datasets problematise the application of prediction models in practice. Representation bias is a prevalent form of bias found in the majority of datasets. This bias arises when training data inadequately represents certain segments of the data space, resulting in poor generalisation of prediction models. Despite AI practitioners employing various methods to mitigate representation bias, their effectiveness is often limited due to a lack of thorough domain knowledge. To address this limitation, this paper introduces human-in-the-loop interaction approaches for representation debiasing of generated data involving domain experts. Our work advocates for a controlled data generation process involving domain experts to effectively mitigate the effects of representation bias. We argue that domain experts can leverage their expertise to assess how representation bias affects prediction models. Moreover, our interaction approaches can facilitate domain experts in steering data augmentation algorithms to produce debiased augmented data and validate or refine the generated samples to reduce representation bias. We also discuss how these approaches can be leveraged for designing and developing user-centred AI systems to mitigate the impact of representation bias through effective collaboration between domain experts and AI.

AI needs to be fair and robust, especially to meet demands of new regulation. Regular assessments are key but it is unclear how we can involve stakeholders without a background in AI in these efforts. This position paper provides an overview of the problems in this area, discusses the current work and looks ahead to future research needed to make headway in user-centric assessment of AI.

The Internet offers a wealth of content, making it increasingly difficult for users to navigate website information. The volume of hyperlinks on a website often leaves users struggling with content overload, hindering their ability to find relevant information of high interest. This problem highlights the critical need for tools to improve the user experience by providing personalized hyperlink recommendations on a specific website. This paper introduces HypeRec, a browser extension that attempts to address this problem by leveraging and comparing different machine learning and recommendation algorithms to guide users to content consistent with their interests and preferences. Our approach involves extracting hyperlinks from a webpage and subjecting the corresponding textual content to natural language processing techniques. In this way, it simplifies the users’ navigation within a website and promotes a more intuitive and satisfying web browsing experience.

By offering a large number of highly diverse resources, learning platforms have been attracting lots of participants, and the interactions with these systems have generated a vast amount of learning-related data. Their collection, processing and analysis have promoted a significant growth of machine learning and knowledge discovery approaches and have opened up new opportunities for supporting and assessing educational experiences in a data-driven fashion. Being able to understand students’ behavior and devise models able to provide data-driven decisions pertaining to the learning domain is a primary property of learning platforms, aiming at maximizing learning outcomes. However, the use of knowledge discovery in education also raises a range of ethical challenges including transparency, reliability fairness, and inclusiveness. In this workshop event, we focus on providing a common ground for researchers and practitioners working in this vibrant area, with the ultimate ambitious goal of bridging the UMAP community with the domain-oriented educational sister communities.

Educational technologies are being used more and more in secondary school settings. This increases the amount of students’ learning related data produced and stored. To keep up with this rise and to get most out of the collected data, teachers need digital tools that support and facilitate their pedagogical decision-making process. Learning analytics dashboards can be a good source to provide teachers with necessary insights into their students’ learning processes. However, for such tools to be effective and actionable, they have to be aligned with teachers’ needs and thus, provide and visualize data in a concise and structured way. We therefore conducted a survey study with 11 English teachers from K-12 secondary schools in Germany who evaluated the assumed usefulness of possible dashboard features. Based on these findings, we developed a teacher dashboard incorporating the most desired functionalities, such as a quickly accessible summary of strengths, weaknesses and support needs, or an overview of current misconceptions and competencies alongside additional metrics in order to support multiple teaching practices. The implementation and the underlying calculations are described, focusing on the importance of learners’ process data to provide teachers with a detailed and revealing view on their students’ and class learning states. In an evaluation study of the dashboard’s prototype with mock data, teachers (n=6) gave high ratings for the dashboard’s usability.

The metaverse represents a persistent, online 3D universe where people can interact, socialize, and work toward common goals. Education represents a key application domain, as it has the potential to enhance experiential learning and collaboration between learners and between learners and educators. However, challenges to the widespread adoption of educational metaverses persist. This paper focuses on emotional isolation, i.e., the feeling of emotional disconnection or loneliness, which can hinder learners’ motivation and participation. Machine learning-enabled emotional recognition systems have the potential to address this challenge, offering educators with feedback on the emotional states of learners within the metaverse. Yet, the integration of emotion recognition systems raises ethical concerns regarding consent, privacy, and algorithmic bias. In this short paper, we conduct a first step toward extracting ethical considerations from the literature on the use of emotion recognition in the educational metaverse. Then, we report these guidelines and finally implement one of the most critical —i.e., protection of privacy— within SENEM, an educational metaverse platform available in the literature. Through this research, we aim to raise awareness within the research community and promote responsible deployment of emotion recognition technology in educational metaverses, aiming to create a supportive and inclusive learning environment for all students.

In recent years, the integration of artificial intelligence (AI) in education has collected significant attention due to its potential to revolutionize learning experiences and support student skill development. This study delves into the dynamics of student interactions with AI support within the domain of C programming education, with a specific focus on the utilization of ChatGPT, a conversational AI model, during training sessions. Through manual clustering analysis, this research unveils distinct patterns of student engagement, elucidating diverse problem-solving approaches and varying levels of interaction with ChatGPT. Our findings underscore the importance of acknowledging individual differences in learning strategies and preferences, highlighting the necessity for personalized educational interventions tailored to meet the diverse needs of learners. However, despite the strides made in AI-supported learning, gaps persist in the existing literature, particularly concerning our understanding of how students approach prompts and exercises when utilizing AI-driven educational tools. This research aims to address this gap by shedding light on the nuanced dynamics of student-AI interactions during training of C programming, offering insights into effective pedagogical strategies and instructional design principles for integrating AI technologies into educational settings. This study makes a significant contribution to the continuous endeavors of educators and AI developers by furthering the discussion on AI-facilitated learning. It aims to enhance student engagement, learning outcomes, and overall educational experiences through the integration of technology into learning environments.

Due to the large amount of available e-learning data, identifying relevant information from e-learning data presents significant challenges. A recommendation system is a popular solution to provide relevant data to any user but it also faces challenges such as scalability, processing large volumes of data, addressing the cold start problem, predicting personalized recommendations, and providing an adaptive recommendation, etc.. In this paper, we present an efficient knowledge graph-based recommendation framework, which can provide personalized e-learning recommendations to existing or new target learners without having enough historical data of that target learner. The proposed framework includes five modules i.e. Data Module, Knowledge Graph Representation Module, Community Building Module, Graph Embedding Module, and Recommendation Module. It is based on knowledge graphs to deal with huge amounts of data and to identify the hidden relationships between data. The proposed framework aims to provide personalized recommendations to learners, it utilizes a clustering-based community generation model for better identifying the interests or preferences of learners.

Ontologies form the core of knowledge graphs, which act as faithful, semantic-rich sources for training models in delivering explainable recommendations. These models learn to extract logical paths between learners and resources to be recommended within the knowledge graph, according to behavior- and content-based patterns. Extracted paths are then used not only to provide recommendations, but also to generate accompanying textual explanations. Despite the potential of this approach, current ontologies derived from the traditional learner-resource interaction data fall short in terms of richness from an educational perspective. Conversely, general-purpose ontologies, while comprehensive in educational aspects, are overly complex for recommendation tasks. Unfortunately, a suboptimal ontology might prevent to articulate reasoning paths, and thus explanations, relevant for learners within the knowledge graph. To counter this limitation, in this paper, we propose LOXER, a novel ontology designed to unlock learner-centered logical paths for explainable educational recommendation. Our design integrates insights from diverse sources, including feedback from a local co-design group of learners, observations from specialized traditional large-scale educational recommendation datasets, and connections with well-known vocabularies of other existing ontologies. To validate our ontology, we conducted an evaluation of the explanation types it enables, involving university and lifelong learners and assessing explanation properties like effectiveness, decision-making speed, motivation, satisfaction, and confidence. Results show our ontology’s ability to foster diverse considerations during the learners’ decision-making process and to establish a semantic structure for knowledge graphs for explainable recommendation.

This study proposes a method for developing a user knowledge model based on their past learning experiences. The focus is on analyzing academic data, particularly lesson records, to extract information about educational concepts. The ultimate goal is to construct a comprehensive profile that reflects the user’s accumulated knowledge throughout their learning journey. Two distinct methods are introduced for concept extraction: a gazetteer-based Named Entity Recognition approach and prompt engineering using ChatGPT. The effectiveness of these methods is assessed through a case study involving a graduate student at the University of Pisa. These knowledge profiles hold significant relevance in today’s educational landscape. With the prevalence of lifelong learning, individuals from diverse academic backgrounds participate in professional development courses. This diversity in past learning experiences can pose a challenge for instructors and course designers who must adapt lessons to be understandable and engaging for an audience with heterogeneous knowledge bases. The analysis of academic data offers a systematic approach to modeling each individual’s acquired knowledge. This, in turn, facilitates the personalization of learning content and pathways based on students’ unique learning experiences. The outcome is an inclusive learning environment that caters to the specific needs of each participant, thereby promoting compelling and stimulating learning experiences.

Integrating Artificial Intelligence (AI) in educational settings has brought new learning approaches, transforming the practices of both students and educators. Among the various technologies driving this transformation, Large Language Models (LLMs) have emerged as powerful tools for creating educational materials and question answering, but there are still space for new applications. Educators commonly use Multiple-Choice Questions (MCQs) to assess student knowledge, but manually generating these questions is resource-intensive and requires significant time and cognitive effort. In our opinion, LLMs offer a promising solution to these challenges. This paper presents a novel comparative analysis of three widely known LLMs – Llama 2, Mistral, and GPT-3.5 – to explore their potential for creating informative and challenging MCQs. In our approach, we do not rely on the knowledge of the LLM, but we inject the knowledge into the prompt to contrast the hallucinations, giving the educators control over the test’s source text, too. Our experiment involving 21 educators shows that GPT-3.5 generates the most effective MCQs across several known metrics. Additionally, it shows that there is still some reluctance to adopt AI in the educational field. This study sheds light on the potential of LLMs to generate MCQs and improve the educational experience, providing valuable insights for the future.

Learner modeling is pivotal in different applications of adaptive and personalized systems in the educational domain, such as recommender systems and intelligent tutoring systems. However, how learner models are inferred and used in the these systems are often not transparent to learners. In many cases, learner models are presented as a black-box, where learners have no means to control or modify their models. To address these issues, in this paper, we present an innovative approach to learner modeling, particularly focusing on modeling learners’ knowledge states. To this end, we combine Personal Knowledge Graphs (PKGs), Graph Convolutional Networks (GCNs), and transformer sentence encoders (SBERT) to construct a transparent learner model. Specifically, we explicitly involve learners in modeling their knowledge state by enabling them to mark concepts as ’Did Not Understand’ (DNU) in the MOOC platform CourseMapper. This results in the construction of a user-controllable and scrutable PKG for the learner, thus increasing the transparency of the learner modeling process. Furthermore, we leverage GCNs and SBERT to model the learner knowledge state based on an enhanced representation of their DNUs. In this way, we provide a simple yet effective method for learner modeling which can be used to improve performance in downstream tasks, such as adaptive systems, recommendation, and personalized search.

Wearable Devices (WDs), encompassing a spectrum from smartwatches to fitness trackers, continuously furnish a wealth of physiological and activity-related data. This trove of information facilitates the creation of robust user models, offering a dynamic lens into users’ daily lives, health patterns, and interaction behaviours. Furthermore, the integration of Brain-Computer Interfaces (BCIs), directly interfacing with neural signals, presents a distinctive vantage point into cognitive processes and emotional states. This integration enriches user models, providing a profound understanding of mental states and engagement levels. While BCIs cannot be strictly categorized as WDs, the latest hardware developments are reaching a size comparable to earphones, anticipating their wearable integration in the near future. However, the exploitation of data from these devices for user modelling and profiling, enhancing personalized activities such as music listening and movie watching, remains an area ripe for exploration.

This workshop proposes an in-depth exploration of the transformative impact that data from WDs and BCIs can exert on user modelling. This initiative seeks to pave the way for a nuanced comprehension of individual preferences, cognitive states, and overall user experiences. The workshop invites researchers, practitioners, and enthusiasts to the convergence of wearable technology, BCIs, and user modelling. Through interactive sessions and discussions, participants will delve into the methodologies, challenges, and opportunities associated with harnessing data from these innovative sources. By fostering collaboration and facilitating knowledge exchange, the workshop aims to propel the current understanding of user modelling by exploiting WDs and BCIs. Attendees will acquire insights into cutting-edge research findings, practical applications, and potential future developments within this rapidly evolving field. Ultimately, the workshop aspires to inspire new research directions, catalyze interdisciplinary collaborations, and cultivate innovative solutions that leverage the synergy between wearable technologies and BCIs to elevate the field of user modelling to unprecedented heights.

In an era characterized by unprecedented virtual connectivity, paradoxically, individuals often find themselves disconnected from genuine human interactions. The advent of remote working arrangements, compounded by the influence of digital communication platforms, has fostered a sense of isolation among people. Consequently, the prevailing socio-technological landscape has underscored the critical need for innovative solutions to address the emotional void. Conversational systems help people improve their everyday tasks with informative dialogues, and recent applications employ them to target emotional support conversation tasks. Nevertheless, their understanding of human feelings is limited, as they depend solely on information discernible from the text or the users’ emotional declarations. Recently, Brain-Computer Interfaces (BCIs), devices that analyze electroencephalographic (EEG) signals, have increasingly become popular given their minimally invasive nature and low cost, besides enabling the detection of users’ emotional states reliably. Hence, we propose ARIEL, an emotionAl suppoRt bcI dEvices and Llm-based conversational agent that aims at supporting users’ emotional states through conversations and monitoring them via BCI. In this way, it is possible to comprehend the users’ feelings reliably, thus making the conversational agent aware of users’ emotional evolution during conversations. Our framework makes the LlaMA 2 chat model communicate with an emotion recognition BCI-based system to achieve the emotional support conversation goal. Also, we present a controlled running example that shows the potential of our model and its effective functioning, made possible by a wisely designed hard-prompt strategy. In the future, we will conduct an in-vivo experiment to evaluate the system and its components.

Objectives: This research delves into the dynamics of human-robot interaction (HRI) in retail environments, with a focus on robot detection from videos captured via an eye-tracking system. Methods: The study employs YOLOv8-nano model for real-time robot detection during grocery shopping tasks. All videos were processed using the YOLOv8 model to test inference speed while performing eye-tracking data analysis as a case study. Results: The YOLOv8 model demonstrated high precision in robot detection, with a mean average precision (mAP) of approximately 97.3% for Intersection over Union (IoU), 100% precision, and 99.87% recall for box detection. The model’s ability to process an average of 160.36 frames per second (FPS) confirmed its suitability for real-time applications. In the case study on the impact of a robot’s presence on human eye movements, the presence of a robot contributes to greater consistency in gaze fixation behavior, potentially leading to more predictable patterns of visual attention. Conclusion: The study’s findings contribute significantly to the design of safer and more efficient cobot systems. They provide a deeper understanding of human responses in real-world scenarios, which is crucial for the development of effective HRI systems.

In electroacoustic music composition, particularly in sound synthesis techniques, Deep Learning (DL) provides very effective solutions. However, these architectures generally have a high level of automation and use textual language for human interaction. To improve the relationship between composers and artificial intelligence systems, brain-computer interfaces (BCIs) are an effective and direct systems, which have led to considerable improvements in this area. The proposed system employs emotion recognition through electroencephalogram (EEG) signals to control four Variational Autoencoders (VAE) that generate new sound textures. A dataset was acquired using the MUSE2 headset to train four Machine Learning (ML) models capable of classifying human emotions based on Russell’s circumplex model. VAEs were trained to produce different sound variations from an audio dataset that allows composers to integrate their sounds. In addition, a graphical user interface (GUI) was developed to facilitate the real-time generation of sound textures, with the support of an external MIDI controller. This GUI continuously provides visual information about the detected emotions and the activity of the left and right brain hemispheres.

In recent years, data-driven methodologies based on artificial intelligence (AI) technologies are taking the lead in many biomedical fields, including the interpretation of electroencephalogram (EEG) experiments collected by wereable Brain Computer Interfaces (BCI). Yet, the effectiveness of data in medical research hinges significantly on its volume and diversity. Traditional biomedical research faces constraints due to limited access to datasets often confined within individual medical centers, hindering potential breakthroughs. Reluctance to share patient data across institutions is driven by ethical, legal, and privacy concerns, with laws like GDPR rigorously protecting against privacy breaches. Overcoming these obstacles requires a paradigm shift towards decentralized AI training procedures, enabling secure and efficient usage of sensible data. In this work, we examine and evaluate the potential of federated learning for the task of emotion recognition from BCI data, focusing on performance with respect to centralized approaches. Our focus lies on comparing its performance against centralized approaches, delving into key metrics such as accuracy, efficiency, and privacy preservation.

This study explores the application of Artificial Intelligence (AI) in the diagnosis of Mild Cognitive Impairment (MCI) and Alzheimer’s Disease (AD), through Human-Computer Interaction (HCI), Human-Centered AI (HCAI), and Explainable AI (XAI). It evaluates three user interfaces designed to integrate AI insights with the clinical understanding of neurologists, aiming to refine diagnostic processes. Neurology professionals were involved to gauge their knowledge and confidence in the AI-supported diagnoses. Utilizing a remotely administered questionnaire, this research investigates clinicians’ views on XAI outputs, focusing on how results are visualized and their ability to engender trust in AI’s clinical utility. This method emphasizes the importance of clear, trustworthy AI systems in healthcare and underscores the essential role of effective human-AI collaboration in enhancing patient care and diagnostic precision.