Recommender systems can produce item recommendations tailored to user preferences and assist user decision making in several real-world applications. Context-aware recommender systems can be built and developed to adapt the recommendations to different contextual situations, since user preferences may vary from contexts to contexts (e.g., time, location, companion, etc.). Recently, the deep learning and neural network techniques have been applied to help build better recommendation models. In this paper, we extend and propose a general neural contextual matrix factorization framework, evaluate and compare a family of these neural contextual matrix factorization models for context-aware recommendations. Particularly, we exploit and analyze the impact on the performance of context-aware recommendations by considering two factors – the component(s) where contexts can be fused into, and the embedding mode utilized to represent context situations.

In this paper, we present a strategy to introduce natural language preference elicitation in a virtual assistant for the movie domain. Our approach allows users to express preferences on objective movie features (e.g., actors, directors, etc.) that are extracted from a structured knowledge base, as well as on subjective features that are collected by mining movie reviews. The effectiveness of the approach was evaluated in a user study (N=103), where our strategy was integrated in a virtual assistant that acquires users’ preferences expressed in form of natural language statements and generates a suitable movie recommendation. Results showed that users experience some difficulties in expressing their preferences in terms of subjective features. However, when people succeed in expressing their preferences by also using subjective properties, this generally leads to better recommendations.

Information Retrieval (IR) systems provide users with a magnitude of information. Complex information needs of users result normally in entire online search sessions that can not be reduced to a singular query. During such sessions complex search activities are executed that comprise several aspects of the users search behavior. One crucial aspect is the users reading activity. To advance towards more adaptive IR systems which recognize the desired interests, we focus on a user model designed for automatic reading detection. This might serve as a measurement for the user investment towards a particular web-page, which should correlate with its relevance. In this work, we propose an entire pipeline from data representation to model prediction. We comparatively evaluate 9 models for automatic reading detection to achieve an accuracy of 79.23% (or 20.77% error-rate). By using such models, we argue that we are able to analyze aspects of the users search behavior and to draw conclusions about their underlying search activity.

Context-aware recommender systems were proposed and built to adapt the recommendations to different context situations. With the development of deep learning based recommendation techniques, the neural network models have also been utilized to improve the quality of the context-aware recommendations. Recently, we released the DeepCARSKit library which is an open-source and deep learning based library for context-aware recommendations. It provides a unified platform for implementing and evaluating context-aware recommendation models based on neural networks. This paper provides a short summary of the DeepCARSKit library, and delivers a user guide to help better use and evaluate the library.

Due to domain-specific challenges such as short item lifetimes and continuous cold-start issues, news recommender systems rely more on content-based methods to deduce reliable user models and make personalized recommendations. Research has shown that alongside the content of an item, the way it is presented to the users also plays a critical role. In this work, we focus on the effect of incorporating expressed emotions within news articles on recommendation performance. We propose a neural news recommendation model that disentangles semantic and emotional modeling of news articles and users. While we exploit the textual content for the semantic representation, we extract and combine emotions of different information levels for the emotional representation. Offline experiments on a real-world dataset show that our approach outperforms non-emotion-aware solutions significantly. Finally, we provide a future outline, where we plan to investigate a) the online performance and b) the explainability/explorability of our approach.

Even though the Internet and social media are usually safe and enjoyable, communication through social media also bears risks. For more than ten years, there have been concerns regarding the manipulation of public opinion through the social Web. In particular, misinformation spreading has proven effective in influencing people, their beliefs and behaviors, from swaying opinions on elections to having direct consequences on health during the COVID-19 pandemic. Most techniques in the literature focus on identifying the individual pieces of misinformation or fake news based on a set of stylistic, content-derived features, user profiles or sharing statistics. Recently, those methods have been extended to identify spreaders. However, they are not enough to effectively detect either fake content or the users spreading it. In this context, this paper presents an initial proof of concept of a deep learning model for identifying fake news spreaders in social media, focusing not only on the characteristics of the shared content but also on user interactions and the resulting content propagation tree structures. Although preliminary, an experimental evaluation over COVID-related data showed promising results, significantly outperforming other alternatives in the literature.

Recommender systems have recently been criticized for promoting bias and trapping users into filter bubbles. This phenomenon not only limits potential user interactions but also threatens the broadness of content consumption. In a music recommender, for example, this situation can limit user perspective as music allows people to develop cultural knowledge and empathy. As a fundamental characteristic of users’ content consumption is its diversity, it is necessary to break the bubbles and recommend potentially relevant and diverse songs from outside the influence of such bubbles. To address this problem, we present MRecuri (Music RECommender for filter bUbble diveRsIfication), a music recommendation technique to foster the diversity and novelty of recommendations. A preliminary evaluation over Last.fm listening data showed the potential of MRecuri to increase the diversity and novelty of recommendations compared with state-of-the-art techniques.

Sport climbing has recently gained large popularity among tourists as a recreational activity. Many people are interested to climb the most beautiful rock climbing places around the world. This has pushed the creation of a large number of climbing routes, to accommodate more and more enthusiasts. However, climbers are not facilitated in their search of routes to climb with any advanced tool, especially in outdoor climbing: they are only provided with either printed or electronic guidebooks, which cannot generate recommendations based on the user’s preferences. Well-tailored climbing routes recommendations have a potential interest for all the involved stakeholders: the users and the companies providing the route information in the form of websites, or guidebooks. To this end, we propose a Content-based Climbing Recommender System prototype. An initial usability study based on the Software Usability Scale (SUS) proves the first version of the prototype to be well-designed (obtained SUS score of 71.6), and the updated version of a system addressing usability problems received an excellent evaluation score (SUS score is 89.3).

This paper explores the issues of optimizing the user experience in a complex, VR-based anti-bullying educational environment. The prototype software is aimed at training early school aged children and uses natural language processing (NLP) to create a more immersive and realistic educational user experience for different gender and age groups.

We introduce a novel system for personalized recommendations to conference attendees, to highlight the papers in the program that best match the attendee’s interests. To this end, we extend traditional structure-agnostic recommender system techniques through the use of deep learning, to exploit the rich semantic and topological information given by the abstracts of the papers and the citation relationship. The ultimate goal is twofold: i) to help attendees single out from a rich program the papers they most likely would like to see presented, and ii) to perform a tailored advertisement of an upcoming event to past attendees by catching their attention with specific contributions in the program of the conference.

Robots will eventually enter our daily lives and assist with a variety of tasks. Especially in the household domain, robots may become indispensable helpers by overtaking tedious tasks, e.g. keeping the place tidy. Their effectiveness and efficiency, however, depend on their ability to adapt to our needs, routines, and personal characteristics. Otherwise, they may not be accepted and trusted in our private domain. For enabling adaptation, the interaction between a human and a robot needs to be personalised. Therefore, the robot needs to collect personal information from the user. However, it is unclear how such sensitive data can be collected in an understandable way without losing a user’s trust in the system. In this paper, we present a conversational approach for explicitly collecting personal user information using natural dialogue. For creating a sound interactive personalisation, we have developed an engaging dialogue strategy. In an online study, the engaging strategy was compared to a baseline dialogue strategy for interactive personalization. Overall, using dialogue for interactive personalization has generally shown positive user reception.

Deep generative models have emerged as one of the most actively researched topics in artificial intelligence. An area that draws increasing attention is the automatic generation of music, with various applications including systems that support and inspire the process of music composition. For these assistive systems, in order to be successful and accepted by users, it is imperative to give the user agency and express their personal style in the process of composition.

In this paper, we demonstrate ReStyle-MusicVAE, a system for human-AI co-creation in music composition. More specifically, ReStyle-MusicVAE combines the automatic melody generation and variation approach of MusicVAE and adds semantic control dimensions to further steer the process. To this end, expert-annotated melody lines created for music production are used to define stylistic anchors, which serve as semantic references for interpolation. We present an easy-to-use web app built on top of the Magenta.js JavaScript library and pre-trained MusicVAE checkpoints.

Data-driven persona generation can benefit from stakeholder inputs while offloading the complexities of high-dimensional datasets. To this end, we present Survey2Persona (S2P), an interactive web interface for real-time persona generation from survey data. The users of the web interface—the designers—can upload survey data and have the interface automatically generate personas. Researchers and practitioners can use S2P to explore different respondent types in their survey datasets in a privacy-preserving manner, which is akin to making the analytical journey more productive, enjoyable, and human-centered. We make the system publicly available and provide argumentation about its novelty and value for user modeling and human-computer interaction communities.

Today, online shops offer a variety of components to support users in finding suitable items, ranging from filters and recommendations to conversational advisors and natural language chatbots. All these methods differ in terms of cognitive load and interaction effort, and, in particular, in their suitability for the specific user. However, it is often difficult for users to determine which method to use to reach their goal. Moreover, as the settings are not propagated between the methods, there is a lack of support for switching components. In this paper, we study the reasons for using the different components in more detail and present an initial proposal for a multi-method approach that provides a more seamless experience, allowing users to freely and flexibly choose from all available methods at any time.

American local newspapers have been experiencing a large loss of reader retention and business within the past 15 years due to the proliferation of online news sources and social media. This has led to a disturbing trend where local journalism and local news outlets are being forced out of business often leaving whole communities without a key source of credible information. This trend has a potentially broad societal impact as these key anchors of local trust and democracy are slowly becoming extinct. Local media companies are starting to shift from an advertising-supported business model to one based on subscriptions to mitigate their financial crises. But with strong competition from a variety of online news sources, these companies need to increase user engagement by providing significant added value. Providing more personalized content in the local context may be one way that these companies can succeed in this effort. Recommender system technologies are the primary enabling mechanisms for delivering such personalized content. However, using standard machine learning models that focus on users’ global preferences is not appropriate in this context because the local preferences of users exhibit some specific characteristics which do not necessarily match their long-term or global preferences in the news. The overall goal of this research is to develop predictive models that more effectively derive user engagement through automatic personalization. Effective recommender systems may be among the tools that can help reverse the current decline of interest in local news sources. Our research explores approaches to learning localized models from user interaction data with news articles, particularly in news categories where there is intense local interest and there is a significant difference between users’ global and local news preferences. Specifically, we propose using such localized models in a session-based recommender system where the system can switch between users’ global and local preference models automatically when warranted. We report experiments performed on a news dataset from a local newspaper show that these local models, particularly the Life-and-Culture news category, do indeed provide more accuracy and effectiveness for personalization which, in turn, may lead to more user engagement with local news content.

Recently, Twitter has become the social network of choice for sharing and spreading information to a multitude of users through posts called ‘tweets’. Users can easily re-share these posts to other users through ‘retweets’, which allow information to cascade to many more users, increasing its outreach. Clearly, being able to know the extent to which a post can be retweeted has great value in advertising, influencing and other such campaigns. In this paper we propose ViralBERT, which can be used to predict the virality of tweets using content- and user-based features. We employ a method of concatenating numerical features such as hashtags and follower numbers to tweet text, and utilise two BERT modules: one for semantic representation of the combined text and numerical features, and another module purely for sentiment analysis of text, as both the information within text and it’s ability to elicit an emotional response play a part in retweet proneness. We collect a dataset of 330k tweets to train ViralBERT and validate the efficacy of our model using baselines from current studies in this field. Our experiments show that our approach outperforms these baselines, with a 13% increase in both F1 Score and Accuracy compared to the best performing baseline method. We then undergo an ablation study to investigate the importance of chosen features, finding that text sentiment and follower counts, and to a lesser extent mentions and following counts, are the strongest features for the model, and that hashtag counts are detrimental to the model.

In the communities of UMAP, RecSys, and similar venues, fairness of recommender systems has primarily been addressed from the perspective of computer science and artificial intelligence, e.g.,  by devising computational bias and fairness metrics or elaborating debiasing algorithms. In contrast, we advocate taking a multiperspective and multidisciplinary viewpoint to complement this technical perspective. This involves considering the variety of stakeholders in the value chain of recommender systems as well as interweaving expertise from various disciplines, in particular, computer science, law, ethics, sociology, and psychology (e.g.,  studying discrepancies between computational metrics of bias and fairness and their actual human perception, and considering the legal and regulatory context recommender systems are embedded in).

Web platforms, such as a video-on-demand services or eCommerce sites, are routinely using Recommender System (RS) to help their users in choosing which item to consume or buy. It is therefore important to understand how the exposure to recommendations can influence the users’ choices and, consequently, the RS’s performance. Important metrics to consider are related to the quality and distribution of the chosen items. This important research focus calls for novel evaluation approaches. A relevant and emerging line of research is based on the simulation of users’ choice behaviour when exposed to recommendations. Simulation-based studies have shown to be useful tools for understanding how an RS performs and its users behave, now and in the future, under various conditions. This paper offers a broad perspective on the field and discusses the potential of simulations in unlocking certain types of analysis that are infeasible by other means. We also discuss the limitations of the current simulation studies.

The Fourth International Workshop on Adaptive and Personalized Privacy and Security (APPS 2022) aims to bring together researchers and practitioners working on diverse topics related to understanding and improving the usability of privacy and security software and systems, by applying user modeling, adaptation and personalization principles. The fourth edition of the workshop includes interdisciplinary contributions from Austria, Belgium, Cyprus, Germany, United Kingdom, and the United States of America, that introduce new and disruptive ideas, suggest novel solutions, and present research results about various aspects (theory, applications, tools) for bringing user modeling, adaptation and personalization principles into privacy and security systems. This summary gives a brief overview of APPS 2022, held both virtually and physically from Barcelona, Spain, in conjunction with the 30th ACM Conference on User Modeling, Adaptation and Personalization (ACM UMAP 2022).

Mobile devices are augmented with advanced computing and sensing capabilities. Due to mobility, they are surrounded by dynamically changing environmental conditions. The immense context information flowing from their environment has paved the way for improved user experience. As users are central to the mobile computing paradigm, context information typically contains sensitive personal data, including user location. Users may need varying levels of privacy at different places and times. Since, as human beings, we are not capable of handling continuously flowing data, it is essential to automatically configure device privacy settings across changing contexts. In this paper, we propose a Context-Adaptive and Personalized Privacy-Preserving System (CAPPPS). It adapts to the changes in location, date, and time as the user context. The system adjusts privacy parameters for mobile device location concerning user privacy requirements while using a location-based system (LBS). We adopt a differential privacy approach to perturb user locations. We evaluated the effectiveness of CAPPPS using actual user trajectories. Results show that CAPPPS is effective for providing adaptive and personalized location privacy based on privacy parameters for changing user context.

The combination of our need for efficient mobility systems coupled with cyber-physical systems has brought about the evolution of Mobility-as-a-Service (MaaS), integrating transport services to provide one-stop access through a custom interface. Our interactions with these MaaS systems lead to a surfeit of data generation and consumption. And for MaaS growth to be sustained, users’ trust in the system, especially in their data privacy, needs to be addressed. In this paper, we use LINDDUN privacy analysis framework to elicit privacy requirements of MaaS systems. We show how User-Dependent Analysis, i.e. modularizing complete use cases to different usage contexts and analysing these usages, can help guide us to discern that usage’s privacy requirements, which can be enacted by relevant MaaS participants.

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 workshop leads to a multidisciplinary research agenda that will inform future research directions and hopefully, forge some research collaborations.

Endowing social robots with the ability to learn and predict the user’s activities during the day is one of the main aims of research in the field of ambient assisted living. Social robots should support older adults with daily activity and, at the same time, they should contribute to emotional wellness by considering affective factors in everyday situations. The main goal of this research is to investigate whether it is possible to learn relations between the user’s affective state and daily routines, made by activities, with the aid of a social robot, Pepper in this case. To this aim, we use the WoMan system able to incrementally learn daily routines and the context in which activities take place. WoMan will be used as a back-end module of the Daily Diary application running on the Pepper robot to collect data concerning daily activities and their relation to emotions and mood. Results of this phase of the research will be used to assess the validity of the approach in ambient assisted living houses for seniors to make the social robot able to provide not only proactive service assistance but also an affective empathic experience.

Recently, it has been considering robotics to face world population aging. According to the WHO, in 2050 there will be about 2.1 billion people over 60 years old worldwide causing a persistent growing need of assistance and a shortage of manpower for delivering congruous assistance. Therefore, seniors’ QoL is continuously threatened. Socially Assistive Robotics proposes itself as a solution. To improve SARs acceptability, it is necessary to tailor the system’s characteristics with respect to the target needs and issues through the analysis of previous and current studies in the HRI field. Through the examination of the state of the art of social robotics in elderly care, past case studies and paper research about SARs’ efficiency, it has been proposed two potential solution examples for two different scenarios, applying two different SARs: Pepper and Nao robots.

Personalization and adaptation are key aspects of designing and developing effective and acceptable social robot tutors. They allow to tailor interactions towards individual needs and preferences, improve engagement and sense of familiarity over time, and facilitate trust between the user and the robot. To foster the development of autonomous adaptive social robots, we present a tutoring framework that recognizes new or previously met pupils and adapts the training experience through feedback about real-time performance and the tailoring of exercises and interaction based on users’ past encounters. The framework is suitable for multiparty scenarios, allowing for deployment in real-world tutoring contexts unfolding in groups.

A preliminary evaluation of the framework during pilot studies and demonstration events in yoga-based training and game scenarios showed that our framework could be adapted to different contexts and populations, including children and adolescents. The robot’s ability to recognize people and personalize its behavior based on the performance of previous sessions was appreciated by participants, who reported the feeling of being followed and cared for by the robot. Overall, the framework can support autonomous robot-led training by allowing monitoring of both daily performance and improvements over multiple encounters. It also lends itself to further expansion to more complex behaviors, with the organic and modular inclusion of more advanced social capabilities, such as redirecting the robot’s attention to different learners or estimating participant engagement.

In this paper we present an educational robot called Wolly, designed to engage children in an affective and social interaction. Indeed, we are now focusing on its role as an educational and affective robot capable of being controlled by coding instructions and at the same time interacting verbally and affectively with children by recognizing their emotions and remembering their interests, and adapting its behavior accordingly.

This paper describes our proposal for enriching personalized social moments and dialogues between human and robot in the context of the Sugar, Salt & Pepper laboratory. The lab focused on the use of the Pepper robot in a therapeutic context to promote autonomies and functional acquisitions in highly functioning (Asperger) children with autism. This paper is focused on a post-hoc work aimed at improving the robot’s autonomous dialogue strategies. In particular we are integrating the robot’s dialogue with a knowledge base to have the robot able to move and reason on an ontology, and thus enriching its dialogue’s strategies. For instance, the taxonomic structure of the ontology could allow Pepper to drive the focus of the conversation to related topics or to more general or specific topics, and, in general, it could improve its capability to manage the conversation and disambiguate the input from the user.

The ability to personalize behaviors is essential for a robot to develop and maintain a long-lasting bond with a user in human-oriented applications, such as a service domain. Service robots must be capable of deducing what actions would be most desirable and best serve the needs and requirements of any interacting users. However, the personalization of a service robot in real-world human-robot interaction (HRI) requires the development of sophisticated mechanisms for identifying differences within the focused group of users, creating a relative user model representation, and finally, devising the varieties of the robot’s behaviors. In this work, we briefly present the multiple methodologies developed for an autonomous bartender robot to personalize its behaviors upon the customers’ moods, attention behaviors, purchasing preferences, personal preferences for interaction, and previous interaction strategies. We expect that the robot would need to serve and interact with multiple customers at the time, as it usually happens in human bartending scenarios. For this reason, our robot has been endowed with the ability to engage multiple users by alternating its attention between them, and personalizing enjoyable interactions through small talk (e.g., welcoming and conversing about topics of general interest related to recent news).

Adaptive and personalized systems have become pervasive technologies, gradually playing an increasingly important role in our daily lives. Indeed, we are now used to interacting with algorithms that help us in several scenarios, ranging from services that suggest us music or movies to personal assistants able to proactively support us in complex decision-making tasks. As the importance of such technologies in our everyday lives grows, it is fundamental that the internal mechanisms that guide these algorithms are as clear as possible. The workshop aims to provide a forum for discussing problems, challenges, and innovative research approaches in this area by investigating the role of transparency and explainability in the recent methodologies for building user models or developing personalized and adaptive systems.

In this paper, we present a framework that supports providing user-specific explanations of AI systems. This is achieved by proposing a particular approach for modeling a user which enables a decision procedure to reason about how much detail to provide in an explanation. We also clarify the circumstances under which it is best not to provide an explanation at all, as one novel aspect of our design. While transparency of black box AI systems is an important aim for ethical AI, efforts to date are often one-size-fits-all. Our position is that more attention should be paid towards offering explanations that are context-specific, and our model takes an important step forward towards achieving that aim.

Recommender systems have gained widespread adoption in many web applications. Modern internet users experience daily interactions with recommender systems. Consequently, users, through these interactive experiences, have developed an inherent understanding of how recommender systems work, what their objectives are, and how the user might manipulate them. We describe this understanding as the Theory of the Recommender. In this paper, we explore the users’ perception and understanding of the recommender system in an empirical study using a grounded theory methodology. To that end, we draw on the cognitive theory of mind to propose a comprehensive theoretical framework that explains the users’ interpretation of the recommender system’s knowledge, reasoning, motivation, beliefs and attitudes. Our findings, based on individual in-depth interviews, suggest that users possess a sophisticated understanding of the recommender system’s behavior. Identifying the user’s understanding is a necessary step in evaluating their impact and improving recommender systems accordingly. Finally, we discuss the potential implications of such user knowledge on recommendation performance.

In this paper, we shed light on explaining user models for transparent recommendation while considering user personal characteristics. To this end, we developed a transparent Recommendation and Interest Modeling Application (RIMA) that provides interactive, layered explanations of the user model with three levels of detail (basic, intermediate, advanced) to meet the demands of different types of end-users. We conducted a within-subject study (N=31) to investigate the relationship between personal characteristics and the explanation level of detail, and the effects of these two variables on the perception of the explainable recommender system with regard to different explanation goals. Based on the study results, we provided some suggestions to support the effective design of user model explanations for transparent recommendation.

The phenomenon of hate messages on the web is unfortunately in continuous expansion and evolution. Even if the big companies that offer their users a social network service have expressly included in their terms of services rules against hate messages, they are still produced at a huge rate. Therefore, moderators are often employed to monitor these platforms and use their critical skills to decide if the content is offensive or not. Unfortunately, this censorship process is complex and costly in terms of human resources. The system we propose in this work is a system that supports moderators by providing them a set of candidate elements to censor with annexed explanations in natural language. It will then be a task of the human operator to understand if to proceed with the censorship and eventually supply feedback to the result of the classification algorithm to extend its data set of examples and improve its future performances. The proposed system has been designed to merge information coming from data, syntactic tags and a manually annotated lexicon. The messages are then processed through deep learning approaches based on both transformer and deep neural network architecture. The output is consequently supported by an explanation in a human-like form. The model has been evaluated on three state-of-the-art datasets showing excellent effectiveness and clear and understandable explanations.

Online debates allow for large-scale participation by users with different opinions, values, and backgrounds. While this is beneficial for democratic discourse, such debates often tend to be cognitively demanding due to the high quantity and low quality of non-expert contributions. High cognitive demand, in turn, can make users vulnerable to cognitive biases such as confirmation bias, hindering well-informed attitude forming. To facilitate interaction with online debates, counter confirmation bias, and nudge users towards engagement with online debate, we propose (1) summaries of the arguments made in the debate and (2) personalized persuasive suggestions to motivate users to engage with the debate summaries. We tested the effect of four different versions of the debate display (without summary, with summary and neutral suggestion, with summary and personalized persuasive suggestion, with summary and random persuasive suggestion) on participants’ attitude-opposing argument recall with a preregistered user study (N = 212). The user study results show no evidence for an effect of either the summary or the personalized persuasive suggestions on participants’ attitude-opposing argument recall. Further, we did not observe confirmation bias in participants’ argument recall, regardless of the debate display. We discuss these observations in light of additionally collected exploratory data, which provides some pointers towards possible causes for the lack of significant findings. Motivated by these considerations, we propose two new hypotheses and ideas for improving relevant properties of the study design for follow-up studies.

We report a diary study of the explanations for the recommendations to characterize the social features in these explanations recorded by five participants over two months. The study reveals several social explanation categories (e.g., personal opinions and personal experiences) and their relationship with user contexts (e.g., location, relevant experience) and recommender attributes (e.g., integrity, expertise) illustrated in a network diagram. Specifically, personal opinions and experiences are two prominent social explanations, mainly associated with user contexts (e.g., users’ preferences and users’ experiences) and several recommender attributes (e.g., politeness, benevolence, and experience). Finally, we discuss several design implications for social explanations and anticipate the value of our findings regarding designing personalized social explanations in recommender systems that aim to build rapport with users, such as conversational recommender systems.

Despite the indisputable personal and societal benefits of regular physical activity, a large portion of the population does not follow the recommended guidelines, harming their health and wellness. The World Health Organization has called upon governments, practitioners, and researchers to accelerate action to address the global prevalence of physical inactivity. To this end, an emerging wave of research in ubiquitous computing has been exploring the potential of interactive self-tracking technology in encouraging positive health behavior change. Numerous findings indicate the benefits of personalization and inclusive design regarding increasing the motivational appeal and overall effectiveness of behavior change systems, with the ultimate goal of empowering and facilitating people to achieve their goals. However, most interventions still adopt a “one-size-fits-all” approach to their design, assuming equal effectiveness for all system features in spite of individual and collective user differences. To this end, we analyze a corpus of 12 years of research in self-tracking technology for health behavior change, focusing on physical activity, to identify those design elements that have proven most effective in inciting desirable behavior across diverse population segments. We then provide actionable recommendations for designing and evaluating behavior change self-tracking technology based on age, gender, occupation, fitness, and health condition. Finally, we engage in a critical commentary on the diversity of the domain and discuss ethical concerns surrounding tailored interventions and directions for moving forward.

Modern societies face many challenges, one of them is the rise of affective polarization over the last 4 decades. In an attempt to understand its reasons, many researchers have questioned the role of Social Media in general, and Recommender Systems (RS) in particular, on the emergence of these extreme behaviors. Diversity in News Recommender Systems (NRS) was quickly perceived as a major issue for the preservation of a healthy democratic debate. However, after more than 15 years of research in Artificial Intelligence on the subject, the understanding of the real impact of diversity in recommendations remains limited. Through a case analysis on the well-known MIND dataset, we propose a critique of the diversity-aware recommendation and evaluation approaches, and provide some take-home messages related to the need of adapted datasets, diversity metrics and analytical methodologies.

The pursuit of algorithmic fairness requires that we think differently about the idea of the “user” in personalized systems, such as recommender systems. The conventional definition of the user in such systems focuses on the receiver of recommendations, the individual to whom a particular personalization output is directed. Fairness, especially provider-side fairness, requires that we consider a broader array of system users and stakeholders, whose needs, interests and preferences may need to be modeled. In this paper, we describe a framework in which the interests of providers and other stakeholders are represented as agents. These agents participate in the production of recommendations through a two-stage social choice mechanism. This approach has the benefit of being able to represent a wide variety of fairness concepts and to extend to multiple fairness concerns.

Fairness-aware machine learning (ML) technology has been developed to remove discriminatory bias, e.g., bias on race and gender. However, there are trade-offs between the metrics of accuracy and fairness in ML models, and different stakeholders prioritize these metrics differently. Hence, to form an agreement on prioritization, workshop approaches encouraging dialogue among stakeholders have been explored. However, it is practically difficult for multiple stakeholders to have conversations at the same place and time. We examined a method of extracting the prioritization of several stakeholders regarding certain metrics using an online survey. We randomly divided 739 crowdsourced participants into 4 stakeholder groups and asked them to rank 5 randomly selected ML models in terms of their metric prioritization. Through this survey, we calculated the prioritization of metrics of each stakeholder group and whether the information on three other stakeholders affects another stakeholder’s prioritization of metrics. With our method, the prioritization of each stakeholder successfully met the requirements of their role. However, metric prioritization is not affected by information on the other stakeholders. Furthermore, demographics and attitudes towards decision making scenarios affect each stakeholder’s metric prioritization differently.

Since recommender systems (“RSs”) are used in multiple domains and applications, issues of possible biases and discrimination have become paramount and present a technical challenge. Indeed, fairness in and for RSs is specific as it does not only concern protected attributes but also encompasses notions of user representativeness and item diversity. In short, RSs require multi-stakeholder fairness. However, RSs are generally built on large (and possibly very sparse) datasets, thus precluding the use of very complex debiasing techniques.

Our approach introduces a fairness functional that minimises the loss disparity across groups and avoids a post-processing step. This enables us to adapt most algorithms underlying recommender systems, such as factorisation machine and its many generalisations and debias them. We present an example of such a functional and show that its properties are ideally suited to the case of multi-stakeholder fair RS.

Finally, we demonstrate that our approach works well in practice on benchmark datasets and that partial debiasing is essential, as full debiasing may lead to poor generalisation.

Targeted advertising is the practice of monitoring people’s online behavior and using the collected information to show people individually targeted advertisements. The term (political) microtargeting is often used when the content of those advertisements is political. Some argue that current regulations are limiting businesses, while others argue that the current legal framework does not do enough to protect individuals. However, the people’s voice is mostly neglected within this debate. In this paper, we present a study in which we assess people’s perception and acceptance of targeted advertising in a commercially versus politically oriented context. The results showed that significantly more people are tolerant towards targeted advertising in a commercial setting than to targeted advertising in a political setting. However, people found the political targeted advertisements to be more useful to them to meet their needs than their commercial counterpart. The results confirm and detail the need for regulations regarding a required level of transparency.

The original filter bubble thesis states that the use of personalization algorithms results in a unique universe of information for each of us, with far-reaching individual and societal consequences. The ambiguity of the original thesis has prompted both a conceptual debate regarding its definition and has forced empirical researchers to consider their own interpretations. This has led to contrasting empirical results and minimal generalizability across studies. To reliably answer the question of whether filter bubbles exists, on what platforms, and what caused them, we need a systematically and empirically verifiable definition of the filter bubble that can be used to develop rigorous tests for the existence and strength of a filter bubble. In this paper, we propose an operationalized definition of the (technological) filter bubble and interpret previous empirical work in light of this new definition.

Group modeling adaptation and personalization is an area explored in parallel by two different research communities. On the one hand, the user modeling community focuses on the preferences aggregation problem: how to combine preferences of individuals in a group so as to personalize, adapt, and explain content for this group to consume or experience? On the other hand, the computer-supported collaboration community focuses on the group formation problem: how to construct a group that will work together efficiently to solve a particular task? This area becomes increasingly significant as work becomes more flexible, online, and distributed. The connecting tissue between both communities is the urgent need to design algorithms, whether for recommending group content or group formations, that steer away from top-down algorithmic decision-making, which has proven to stifle user agency and create power inequalities between users and algorithms. The aim of the workshop is, for the first time, to bring together the two communities working on the two sides of Group Recommendations, with an overall goal to rethink group recommendation and shift paradigms from the current algorithm-centric to a user- and group-centric focus.

Forming teams of learners is a task that presents numerous challenges for educators increasingly relying on automated tools to optimize the process. The problem increases in difficulty in online classroom settings, where educators have little familiarity with the students. In this work, we present a User as Wizard study where 108 online crowd participants formed four teams of three teammates each from a pool of twelve dummy learner profiles. The profiles contained information about the learners’ Conscientiousness, Openness, and cognitive ability levels. These attributes were derived from a pre-study with a smaller sample of crowd participants (N=52) rating the relevance of the Big Five personality traits and cognitive ability in team formation for educational purposes. The User as Wizard study shows that most people tend to form within (meaning most attributes of the teammates even out) and between (meaning the teams have similar attributes averages) balanced teams. It also shows that people perceive Conscientiousness and Openness as two of the most relevant personality traits when profiling learners for team formation. We compare these results to the probability of them being random and discuss the findings in the light of human-centered modeling of system designs and automation in education.

In recent years, several recommendation systems have been introduced to improve the user experience of players in video games. In Multiplayer Online Battle Arena (MOBA) games, a popular game genre, these systems are useful for recommending items for a character during a match. Current approaches focus on recommending a fixed set of items based on a character and the other participants. However, a MOBA match is an inherently sequential process, where its past decisions define the current ones. Therefore, it would be desirable to obtain a contextual recommendation considering a specific situation and the previously consumed items. To fill this gap, in this work, we propose HT4Rec for group-aware sequential item recommendation. It consists of a contextual encoder that generates a character-item representation contextualized by the other participants involved in the game, followed by a sequential encoder that captures sequential patterns of the data to recommend the next item. In this way, HT4Rec provides a flexible and unified attention-based network structure to capture both general and long-term preferences. Our evaluations on a Dota 2 video game dataset demonstrate that HT4Rec outperforms well-known sequential recommendation methods on various evaluation metrics. Additional experiments unveil the most important parts of our model and the most relevant inputs according to the attention mechanism, which could be used to interpret the suggested items. Furthermore, we demonstrate that HT4Rec could be applied to a different scenario (movie recommendation) than MOBA games, showing better results than the baseline models.

Group recommendations are a specific case of recommender systems (RS), where instead of recommending for each individual independently, shared recommendations are produced for groups of users. Usually, group recommendation techniques (i.e., group aggregators) are built on top of common ”single-user” RS and the resulting group recommendation should reflect both the overall utility of the recommendation as well as fairness among the utilities of individual group members.

Off-line evaluations of group recommendations were so far resolved either as a tightly coupled pair with the underlying RS or in a decoupled fashion. In the latter case, the relevance scores estimated by underlying RS serves as a ground truth for the evaluation of group aggregators. Both coupled and decoupled evaluation may suffer from different biases that provide illicit advantages to some classes of group recommending strategies.

In this paper, we focus on the decoupled evaluation protocol and possible polarity bias of the underlying RS. We define polarity bias as situations when RS either locally or globally under-estimate or over-estimate the true user preferences. We propose several polarity de-biasing strategies and in the experimental part, we focus on the capability of group aggregation strategies to cope with the polarity biased input data.

Going to restaurants is also a social activity; people often go to restaurants with family, friends, or colleagues. However, most restaurant finder systems, such as TripAdvisor, allow users to search for restaurants matching only one user’s preferences. We present here a system GUI aiming at extending such systems to support the organizer of an event in finding a proper restaurant for her group. The organizer is responsible for expressing the group members’ preferences, analyzing the recommendations, and finally selecting a restaurant. We have identified three recommendation techniques (popularity-based, relevance-based, and critiquing-based) to support such a task. These techniques make different assumptions on the amount of information about the group members’ preferences available to the organizer, and they support alternative choice patterns. Moreover, the proposed system supports in the final decision-making stage by: a) indicating the extent to which a recommended restaurant is attractive to each group member, w.r.t. the entered preferences, b) suggesting a good choice for the group, and c) illustrating the similarity of other groups which have previously bookmarked one of the recommended restaurants.

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 seventh edition of HAAPIE includes 5 long papers and 2 short papers.

Persuasive technologies are interactive systems that are designed to influence people to change their attitudes or behaviours. Persuasive technologies have been used successfully in several domains including health to make people exercise more, shopping to make people buy specific products, and social media to make people contribute better content. In the area of sustainability, its use is not well documented. To contribute to the use of persuasive technologies in sustainability, this paper carries out a literature review of published articles in the area in the past five years and summarizes the main findings based on three main themes: the design and development of the technology to make it adaptive to users, the evaluation of the technology, and the findings from the evaluation. Our results suggest that most persuasive technologies are developed as mobile applications, IoT devices or serious games and the most common behaviour change targeted by the persuasive technologies in this domain are energy conservation and sustainable food management. The most common persuasive strategies that are used are rewards, suggestions and self-monitoring. In terms of evaluation, a self-reported evaluation method was applied by most authors. While the range of evaluation of the developed persuasive technologies was between one hour and one year, the number of recruited participants ranged from two to over nine hundred. The findings from the evaluation were mostly mixed with several authors reporting positive results (behaviour change) for some participants. Based on these results, we suggest considerations for the development of future persuasive technologies for sustainability.

Many people with dementia residing in long-term care may face barriers in accessing experiences beyond their physical premises; this may be due to location, mobility constraints, legal act and/or mental health restrictions. Previous research has suggested that institutionalization increases the co-existing symptoms of dementia, such as aggression, depression, apathy, lack of motivation and loss of interest in oneself and others. Despite the importance of supporting the mental well-being of people with dementia, in many cases, it remains undertreated. In recent years, there has been growing research interest in designing non-pharmacological interventions aiming to improve the Health-Related Quality of Life for people with dementia within long-term care. With computer technology and especially Virtual Reality offering endless opportunities for mental support, we must consider how Virtual Reality for people with dementia can be sensitively designed to provide comfortable, enriching out-world experiences. Working closely with 24 dementia patients and 51 medical and paramedical personnel, we co-designed an intelligent and personalized Virtual Reality system to enhance symptom management of dementia patients residing in long-term care. Through this paper, we thoroughly explain the screening process and analysis we run to identify which environments patients would like to receive as a Virtual Reality intervention to minimize the aforementioned co-existing symptoms of dementia, and the development of an intelligent system using the selected environments, that adapts the content of the Virtual Reality experience based on physiological and eye-tracking data from the patients and their personal preferences.

Monitoring a learner’s performance during practice plays an important role in scaffolding. It helps with scheduling suitable practice exercises and, by doing so, sustains learner motivation and steady learning progress while moving through the curriculum. In this paper we present our approach for monitoring the learning progress of students learning to play piano with Hidden Markov Models. First, we present and implement the so-called practice modes, practice units that are derived from the original task by reducing its complexity and focusing on one or several relevant task dimensions. Second, for each practice mode, a Hidden Markov Model is trained to predict whether the player is in the Mastered or NonMastered latent state regarding the current task and practice mode.

Physical activity is important to improve an individual’s overall well-being. Digital interventions as they use Virtual Reality (VR) and Augmented Reality (AR), have shown success in promoting physical activity (PA) in people of all ages. This work discusses the design of an AR gamified mobile application prototype for promoting physical activity in young adults. The application (app), “The Journey” aims to promote PA in young adults’ users while they explore various touristic sites and also acquire virtual assets. This is achievable at a low cost to users by using smartphones-based AR app to tour any location of interest from the comfort of their home or outdoor. Each user’s step count tracked via mobile device is used to help them navigate the location of interest. The findings from our evaluation of 29 people show that The Journey has the potential to motivate people to improve their PA both indoors and outdoors.

As a result of the COVID-19 pandemic, a lot of people have been forced to work from home. Particularly during video conferences, workers basically invite their colleagues, co-workers and supervisors into their homes, sacrificing portions of their privacy in the process. In this paper, we investigate which home-related and work-related factors are perceived as relevant for privacy. We asked participants to indicate their preferences for videoconferencing settings in various scenarios and also asked about their personal experiences. The results show that power distance plays a role, but that group size and familiarity with other group members are more decisive factors. We discuss implications of our findings in terms of user awareness and the benefits of different context-based default settings for videoconferencing privacy settings.

American local newspapers have been experiencing a large loss of reader retention and business within the past 15 years due to the proliferation of online news sources. Local media companies are starting to shift from an advertising-supported business model to one based on subscriptions to mitigate this problem. With this subscription model, there is a need to increase user engagement and personalization, and recommender systems are one way for these news companies to accomplish this goal. However, using standard modeling approaches that focus on users’ global preferences is not appropriate in this context because the local preferences of users exhibit some specific characteristics which do not necessarily match their long-term or global preferences in the news. Our research explores a localized session-based recommendation approach, using recommendations based on local news articles and articles pertaining to the different local news categories. Experiments performed on a news dataset from a local newspaper show that these local models, particularly certain categories of items, do indeed provide more accuracy and effectiveness for personalization which, in turn, may lead to more user engagement with local news content.

Content-based recommender (CBR) systems take advantage of item characteristics and user propensities for these characteristics in order to select the items that are better suited for a user. Related work has shown that the characteristics of hedonia (pleasure) and eudaimonia (deeper meaning) account for user preferences in the domain of movies. However, labeling items with hedonic/eudaimonic properties and measuring user propensity for eudaimonic/hedonic experiences could be done only through questionnaires. In this work we present the results of our work-in-progress on the prediction of user propensities for eudaimonic and hedonic experiences from a movie preferences dataset. Our results indicate that a range of classifiers that use ratings of movies as features perform substantially better than the average baseline.

ACM PATCH 2022, the 13th International Workshop on Personalized Access to Cultural Heritage, is organized in conjunction with the 30th International Conference on User Modeling, Adaptation and Personalization. It is the meeting point between researchers and practitioners working on personalization in cultural heritage, aiming to enhance the user experience in digital and physical Cultural Heritage sites. The PATCH workshops started in 2007 and they are typically held in conjunction with UMAP, IUI and recently AVI Conference series. This paper summarizes the main ideas addressed in the papers accepted for presentation at PATCH 2022 and for publication in the workshop proceedings.

The paper presents a combined approach to knowledge-based emotion attribution and classification of cultural items employed in the H2020 project SPICE. In particular, we show a preliminary experimentation conducted on a selection of items contributed by the GAM Museum in Turin (Galleria di Arte Moderna), pointing out how different language-based approaches to emotion categorization (used in the systems Sophia and DEGARI respectively) can be powerfully combined to cope with both coverage and extended affective attributions. Interestingly, both approaches are based on an ontology of the Plutchik’s theory of emotions.

Information system department, University of Haifa, Israel, ishimshoni@is.haifa.ac.il

With the growth of the commercial interest in indoor Location-based Services (ILBS), a lot of effort was put into the development of indoor positioning systems. One way to track users in indoor environment is using image-based localization. The user captures an image in front of a desirable place and the system locates him. Such systems use image matching algorithms, and usually they try to match the current image that was captured by the user with all the images that exist in the dataset. The big challenge is the dataset preparation; it has to be representative and minimal as much as it can be to reduce the number of comparisons. Previous works used special equipment to map or scan the environments, and others used human operators who have expertise in image matching algorithms. In this work, we present ARIDF, an automated method that finds a minimal and representative dataset for image-based localization. The human operators should not have any previous knowledge and experience about image matching algorithms.

This document describes the rationale, the implementation and a preliminary evaluation of a semantic reasoning tool developed in the EU H2020 SPICE project to enhance the diversity of perspectives experienced by museum visitors. The tool, called DEGARI 2.0 for values, relies on the commonsense reasoning framework , and exploits an ontological model formalizing the Haidt’s theory of moral values to associate museum items with combined values and emotions. Within a museum exhibition, this tool can suggest cultural items that are associated not only with the values of already experienced or preferred objects, but also with novel items with different value stances, opening the visit experience to more inclusive interpretations of cultural content. The system has been preliminarily tested, in the context of the SPICE project, on the collection of the Hecht Museum of Haifa.

Most Cultural Heritage mobile guides are developed using a location-based hypertext model that guides the exploration of individual itineraries. However, Cultural Heritage places are often immersed in parallel and interlaced stories about art, history and science, which might be relevant to the tourist. Therefore, we are interested in investigating the semantic connections between the narratives that involve different Point of Interests to provide tourists with interconnected views of the Cultural Heritage of a place.

As a first step in this direction, we present the Triangolazioni mobile guide, which allows users to thematically explore places through the navigation of narratives, and to connect narratives to each other based on topic similarity. Triangolazioni can be accessed from mobile phone, tablet and desktop, and supports the physical and virtual exploration of Points of Interest in the area around Torino, Italy.

This paper presents a visualization model supporting the map-based presentation of environmental data about geographic areas. Our model is based on the evaluation of the ecological aspects of a landscape and on the dynamic generation of a layer that shows the levels of naturality, the land use, and other similar characteristics, of the area in focus. We propose this model as an extension of the information that can be provided by a mobile guide with the aim to support the exploration of Natural Heritage by making users aware of the richness and weaknesses of the places they visit and raising the attention towards biodiversity and environment preservation, which are key to green tourism.

In this paper, we present the rationale and the ideas behind META4RS, a museum itinerary recommender system. The system leverages deep learning techniques to acquire data about the visitor’s position while ensuring her anonymity. Moreover, the visitor’s appraisal of the artwork she observes is inferred implicitly based on the emotional reactions she expresses while watching a given artwork. We are not aware of any such recommender system proposed in the research literature. However, this system should ensure several advantages: (i) it is non-intrusive since it makes use of simple badges and off-the-shelf cameras while ensuring the anonymity of the visitor; (ii) it is independent of the type of museum; (iii) it offers personalized itineraries to visitors based on their implicitly inferred interests and preferences. Specifically, we illustrate the background and describe the architecture of the proposed system, discussing the steps required for its implementation. We also provide details of what has already been done and what remains to be done, outlining the open problems.