It is our great pleasure to welcome you to the UMAP 2018 Workshops and Tutorials. In the 26th edition of the ACM Conference on User Modeling, Adaptation, and Personalization there are four workshops and three tutorials.
The FairUMAP Workshop at UMAP 2018 brought together researchers working at the intersection of user modeling, adaptation, and personalization on the one hand, and bias and fairness in machine learning on the other hand.
Recommender systems for news articles on social media select and filter content through automatic personalization. As a result, users are often unaware of opposing points of view, leading to informational blindspots and potentially polarized opinions. They may be aware of a topic, but only be exposed to one viewpoint on this topic. However, recommender systems have just as much potential to help users find a plurality of viewpoints. In this spirit, this paper introduces an approach to automatically identifying content that represents a wider range of opinions on a given topic. Our offline results show positive results for our distance measure with regard to diversification on topic and channel. However, our user study results confirm that user acceptance of this diversification also needs to be addressed in tandem to enable a complete solution.
Public radio broadcasters do not consider personalization in their compliance with their public-service remit so far. However, personalization brings along the risk of filter bubbles, which contradicts with the ideas of the public-service remit. We shed light on the interconnection of personalization and the public-service remit of broadcasters anchored in an analysis of the interstate treaty on broadcasting and tele-media. The contribution of this paper is two-fold. First, we propose an approach how to selectively avoid filter bubbles in personalized radio consumption. Second, we develop a framework that helps to assess the compliance of personalized radio offers with public-service remits.
The trade-off between relevance and fairness in personalized recommendations has been explored in recent works, with the goal of minimizing learned discrimination towards certain demographics while still producing relevant results. We present a fairness-aware variation of the Maximal Marginal Relevance (MMR) re-ranking method which uses representations of demographic groups computed using a labeled dataset. This method is intended to incorporate fairness with respect to these demographic groups. We perform an experiment on a stock photo dataset and examine the trade-off between relevance and fairness against a well known baseline, MMR, by using human judgment to examine the results of the re-ranking when using different fractions of a labeled dataset, and by performing a quantitative analysis on the ranked results of a set of query images. We show that our proposed method can incorporate fairness in the ranked results while obtaining higher precision than the baseline, while our case study shows that even a limited amount of labeled data can be used to compute the representations to obtain fairness. This method can be used as a post-processing step for recommender systems and search.
Traditional recommender systems suggest items by learning from user preferences, but ignore other stakeholders in the whole system. Actually, not only the receiver of the recommendations, but also other stakeholders may come into play, such as the producers of items or those of the system owners. Reciprocal recommender system in dating or job recommendations is one of these examples. However, we may have to simulate the utilities for each type of the stakeholder due to the utility definitions. In this paper, we perform exploratory analysis on a speed-dating data, where the user expectations are clearly defined. We try to build a multi-dimensional utility framework by utilizing multi-criteria ratings. We further analyze the relationship between the utilities and recommendation performance, and achieve a tradeoff as the optimal solution. Even more, the proposed approach is able to outperform the exiting reciprocal recommendation algorithms in precision, recall and overall utilities. Finally, we derive a promising way to define and optimize utilities to be generalized in other applications or domains.
The Diversity Checker is a tool that aims to make it easier for journalists to author their texts with diversity in mind. To provide helpful hints for them in this respect, it is necessary to define how to quantify diversity so that this can be programmed into the tool. At this early stage in the development of the tool, we present a two-fold contribution. First, we offer an analysis on what we mean by "improving diversity". Second, we present the first version of the Diversity Checker, along with some analysis of its current performance.
It is our great pleasure to welcome you to the 3rd International Workshop on Human Aspects in Adaptive and Personalized Interactive Environments (HAAPIE 2018). HAAPIE 2018 (http://haapie.cs.ucy.ac.cy) is a full-day workshop held on 08 July 2018 in conjunction with the 26th ACM Conference on User Modeling, Adaptation and Personalization (UMAP 2018), 08-11 July 2018 in Singapore. 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. HAAPIE embraces the essence of the human-machine co-existence and aims to bring more inclusively the "human-in-the-loop" approach/idea, adequately supporting the rising multi-purpose goals, needs, requirements, activities and interactions of users through new human-centered adaptive and personalized interactive environments, algorithms and systems. It brings together experts, researchers, students and practitioners from different disciplines in order to share ideas and experiences, lessons learned, approaches and results that could substantially contribute to the broader UMAP community.
The way people read digital news - as distinct from what news they read - has emerged as a significant concern for research in user modelling and personalisation. Intuitively, some people read the news frequently and broadly whilst others read it occasionally and selectively. It is likely that these differences in news reading behaviour arise in part from differences in peoples' personalities. We report a study that surveyed the digital news reading habits and personality traits of 241 people. We find correlations between most news reading characteristics (e.g., how much time over a day a person reads news) and some personality traits (e.g Openness-to-Experience). The correlations provide a better understanding of the different types of news reading user and why they read news in different ways. They indicate the value of extending user model profiles to include personality traits along with domain specific activity factors.
Personality, as one of the human factors, has been demonstrated as an influential factor in decision making. Particularly, personality traits can be utilized to identify decision leaders and followers in the context of group decision making. In this paper, we propose an approach to learn user roles (i.e., decision leaders and followers) to improve the performance of group recommendations. More specifically, we utilize the binary particle swarm optimization as the method to assign and learn user roles in each group. Our experimental results based on an educational data reveal that the proposed approach is able to improve group recommendations. However, the learned user roles may not present expected characteristics in terms of the personality traits.
We present a novel approach in the design of self-care applications aimed at women in pre-menopause and menopause, which is based on the use fuzzy rule-based systems (FRBS) for the description of women's health status, behavior, and personality traits, as well as the provision of adaptive interventions. Our main goal is to develop a personalized solution that informs users and promotes healthy behavioral changes in them, as many women are not conscious of the health consequences caused by menopause. To this end, we follow just-in-time adaptive interventions (JITAI), a design framework for the rule-based provision of personalized and on-demand interventions through mobile applications, according to user's health status and behavior. However, these health-related concepts are usually vague and difficult to formalize in a well-constrained manner. Therefore, to overcome these issues, our approach seeks to leverage the knowledge of health professional experts and the capacity of fuzzy logic modelling and inference to formally define such vague concepts and rules. Although we discuss this approach in the context of menopause self-care, in which prior user data may be insufficient for alternative techniques, it could also be applied in other self-care applications facing similar challenges.
Nowadays, the role of the computer and the Internet has been upgraded in the lives of people, witnessing a paradigm shift on how users communicate with each other or with their service providers. Main characteristics of this reality include dynamic strategies and conditions, multivariate random interactions with instant feedback, and holistic activities with no clear origins across mixed realities. Thereupon, it seems that traditional human-computer interaction methods and adaptivity practices lack the appropriate dynamicity to fully address these aspects and to explicitly offer holistic solutions for enhancing user experience. In this paper, we propose an alternative model and formalization by combining concepts like equilibria and refocusing the core of investigation into the actual single objective interactions that are generated by the interplay of two (or more) participatory ends over a time interval and which cannot be recognized a priori. The resulting interactions consist of dynamic suboptimal adaptivity states of the two that provide an Equilibrium State of Interaction and a motivational engagement.
The use of videos in education has attracted considerable research attention. However, in order to gain the most benefits, learners need to actively engage with videos. It is an important, yet challenging, task to disentangle the relation between engagement with videos and learning, and at the same time to take into account relevant individual differences in order to offer personalised support. In this paper we investigate the question: "Can user characteristics relating to self-regulation, knowledge, and experience be leveraged for predicting user engagement with videos?". Our results show that users' domain knowledge and self-regulation abilities can inform overall engagement prediction (inactive, passive and constructive learners), which makes them useful for adaptation and personalisation.
Wallas suggested a four stages model of creative process: a) preparation, b) incubation, c) illumination, d) verification, that has been widely used through the years in several disciplines. In this work we are aiming at defining pattern detection algorithms for modelling the creative process of a user based on the user's activity in MineTest. A qualitative user study allowed us to define and refine patterns related to the creative process of the user while executing a creative task in the game. In addition, through the data collected, important issues have been exposed that will inform future work in the same direction.
Using Personal Assistants (PAs) for tasks in our everyday life is quite common these days. The PAs (Apple Siri, Amazon Alexa, Google Now, Microsoft Cortana etc.) help us to set reminders, find our way through traffic, or send messages to friends and colleagues. While being always available to serve our needs, they collect our personal data to improve their Services and tune their behavior. In order to find out which features (e.g., proactivity, storing of data) are accepted or even desired by potential users to model an adaptive PA with respective spoken output, we conducted an online survey with 1,051 German speaking and 499 US participants. We compared the participants' information on their interaction with different systems, including in-vehicle Spoken Dialog Systems (SDSs), smartphone voice control, and smart home devices (e.g., Amazon Echo, Google Home). We had a look both into output features of spoken dialog by showing dialog examples, as well as posing direct questions on potential system features and its interaction behavior. The results show that especially proactivity is desired by the participants while none of the participants has a positive attitude towards the systems' data usage when asked explicitly.
Advances in technology and virtual reality have put the focus on exergames aimed at the elderly as complementary tools for rehabilitation, due to the beneficial aspects for motor learning they provide. This paper presents a research in progress of an exergame which is being designed for frail older adults. A user centered game design process is being carried out to identify the set of exercises to be included in the exergame.
It is our great pleasure to welcome you to the UMAP 2018 HUM (Holistic User Modeling) Workshop. According to a recent claim by IBM, 90% of the data available today have been created in the last two years. This exponential growth of online information has given new life to research in the area of user modeling and personalization, since information about users' preferences, sentiment and opinions, as well as signals describing their physical and psychological state, can now be obtained by mining data gathered from many heterogeneous sources. We can distinguish two important classes of such data sources. One of these comes from recent trends in Quantified Self (QS) and Personal Informatics, which has emphasized the use of technology to collect personal data on different aspects of people's daily lives. These data can be internal states (such as mood or glucose level) or indicators of performance (such as the kilometers run). The purpose of collecting these data is self-monitoring, performed to gain self-knowledge or to obtain some change or improvement (behavioral, psychological, therapeutic, etc.). Often these data are also exploited for behavior change purposes, for example to increase the user's physical activity. The other key category comes from the enormous amount of textual content that is continuously spread on social networks. This has driven a strong research effort to investigate to what extent such data can be exploited to infer user interests, personality traits, emotions, and knowledge. Moreover, the recent phenomenon of (Linked) Open Data fueled this research line by making available a huge amount of machine-readable textual data that can be used to connect all the data points spread in different data silos under a uniform representation formalism. The main goal of the workshop is to investigate whether techniques for advanced content representation and methodologies for gathering and modeling personal data (e.g. physiological, behavioral) can be exploited to build a new generation of personalized and intelligent systems in domains as diverse as health, learning, behavior change, e-government, smart cities (e.g., by combining mood data and music preferences data to provide recommendations on music to be listened).
Tour planning is a difficult task for those who visit unfamiliar city destinations. Furthermore, building an itinerary becomes more difficult as the number of options, which can be incorporated into travel, increases. The authors aim to propose place of interest (POI) according to the narrative strategy of a tour guide to realize a better personalized mobile tour guide system and establish a method to support efficient route scheduling. As a basic stage, we will herein consider a method of naturally collecting context information of users through an interaction between users and information terminals. In addition, we will introduce a POI recommendation application using the context information being developed.
In this paper we introduce the concept of holistic user profile, intended as a unique representation of a user that merges the heterogeneous footprints she spread on social networks and through personal devices, and we present a framework that supports the creation of such user models. Our holistic user model is based on the insight that each person can be described through different facets, such as her interests, activities, habits, mood, social connections and so on, and each facet can be modeled by gathering and merging information coming from diverse sources. To this end, we designed a platform that automatically acquires personal data coming from both social networks, such as Twitter and Facebook, as well as devices, such as Android smartphones and FitBit wristbands. Next, these rough data are merged, processed and enriched in order to infer high-level features (e.g., user interested in technology, user experiencing a bad mood, sedentary person, and so on) or to extract user behavioral patterns (e.g., places that are frequently visited). Such unique representation of a person, whose strength is the combination of different heterogeneous data points, is finally encoded and stored in our platform and can be made available to both the user itself and to third-party services. In the former case, data are shown through a visual interface. In the latter, holistic user profiles are exposed through this unique entry point and external applications can build new personalized services based on the digital footprints spread by the user.
The growing popularity of e-commerce has ignited the interest of the research community in e-commerce application research and development. For this purpose, variety of applications and resources such as MovieLens and IMDb datasets have been utilized, which present a rich database of items (movies), their attributes and ratings. It is a common practice to utilize both MovieLens and IMDb datasets to produce richer dataset about items and user-generated information such as tags and ratings for conducting research in recommender systems domain. Moreover, the combined dataset contains contextual feature represented by timestamps. However, dealing with a big chunk of the IMDb dataset and its integration with MovieLens dataset remains a challenge due to the various issues such as resource constraints and broken links. To address these issues, a more complex procedure has to be adopted for the integration of both datasets. Hence, this paper introduces iSynchronizer, which is an open-source tool for synchronizing MovieLens and IMDb datasets autonomously with minimum effort. Moreover, the tool includes code to generate a matrix of item features and user preferences over item features from the raw integrated dataset.
This paper presents a personalized interactive map aimed at supporting people with Autism Spectrum Disorder (ASD) in their daily transfers within urban environments. To this end, it aims to model a "complete" representation of the ASD individual by merging a variety of information in a unique user model. Moreover, it exploits crowdsourcing mechanisms to enrich the representation of places that may be considered "safe" by ASD people. As a result, the system is specifically designed for helping them manage stress originated by breakdowns from "spatial routines", by providing recommendations about safe places to reach and giving personalized tools to manage unexpected events.
Websites are more than ever tailoring themselves to their customers, gathering and using the information they are providing in order to offer a differentiated product. Most people are aware of their browser's history and cookies, but with the rise of single-login, geolocation and online profiles, the boundaries are getting blurrier. Companies are collecting data at an exponential rate, jeopardizing their clients' privacy. And, so far, people are making it easy to collect their data since they are so willingly disclosing it online. In addition, the rise of social networks makes the need for privacy protection more crucial than ever. But technology brings new choices, new risks, and new opportunities. In particular, privacy-protection concerns should not hamper the benefits of a society of sharing. Thus, a delicate balance must be reached between these apparently conflicting requirements.
Recent advances in graph and network embeddings have been utilized for the purpose of providing recommendations. Hybrid recommender systems have shown the efficacy of using side information associated with entities. In this work we show how domain specific knowledge can be used to define meta paths within these heterogeneous domains and how these path constrained random walks can be used to embed user preferences in heterogeneous domains. The semantic embeddings generated from heterogeneous knowledge sources combined with user preferences can be used to refine a user's information needs. This representation modeling of users, entities and their associated properties opens up new modalities of interactions for the users to gravitate towards their requirements. In this work we propose the use of semantic embeddings for two kinds of interactive recommendation modalities: 1) exemplar based recommendations 2) "less like this/ more like this" style recommendations. In our opinion providing these modalities would boost the expressive power of exploratory search and recommender systems.
Top-N Recommender Systems usually suffer from intra-list diversity as they are tailored for relevance and predicted rating accuracy. This problem is magnified in the case of cold start setting - resulting in users being restricted to popular set of items and can result in a "rich getting richer eco-system". As a result, in recent years, more attention is being paid to improving the diversity of recommender system results. List creation has become a popular way for users to express preferences over items on online platforms such as imdb.com and goodreads.com. These user curated lists tend to contain a coherent semantic representation of the domain the list of items belong to. List curation can be seen as a way to capture fine grained topic-specific item-lists by users. Understanding and modeling user preferences expressed in these curated lists can help with diverse set of applications such as recommendations, user modeling, session understanding etc. In this paper, we propose an approach to improve the diversity of results generated by Top-N recommender systems, by using Determinantal Point Processes (DPPs) over user curated lists in the movie domain and incorporating them to rerank the Top-N recommender systems. For this work, we use the user curated lists in the imdb.com domain. We evaluate our approach over the Movielens 1-Million dataset and compare the results with other baseline approaches. Our early results show that incorporating semantic similarity expressed in user lists as a diversity proxy results in a more diverse set of recommendations.
E-Learning is transforming the way education is imparted. Today, millions of students take self-paced online courses. However, the content and language complexity often hinders comprehension and this together with lack of immediate help from the course instructor leads to weak learning outcomes. Ability to predict difficult content in real time enables eLearning systems to adapt content as per students' level of learning. The recent introduction of low-cost eye trackers has opened the new class of applications based on eye response. Eye tracking devices can record eye response to the visual element or concept causing a learning difficulty. The response and the variations in eye response to the same concept over time may be indicative of the level of learning. In this paper, we use eye movement measures to predict the levels of learning associated with a term/concept. The main contribution of this study is the spatio-temporal analysis of eye response to a term/concept. Proposed system analyses slide images, extracts words (terms), maps the eye response to words, and prepares a term-response map. A majority voting classifier trained with terms of known learning levels uses this term response map to classify a term as novel or familiar. The proposed system achieves 61% accuracy when predicting learning difficulty.
Mobile banking (MB) services offered by financial institutions are moving from being a strategic advantage to sustain competition fueled by the ubiquity of smartphones among banking customers. Yet, MB apps are difficult to use with small screens and new features being added constantly. While personalization features in the form of recommender systems are commonplace in ecommerce applications, personalization in MB is in early stages. This study focuses on adapting MB application interface by using a data analytic technique; the novelty and contribution of our MB application is its focus on real-time analysis of user's prior interactions with the system and adapts to their needs with real-time analysis of user's prior interactions with the system to improve user experience and usage. PERMA system is a next generation of Adaptive Hypermedia System (AHS) where our MB app provides personalized content, presentation, and navigational support
Smartwatches are becoming more and more popular and offer an alternative to smartphones for frequent but short interactions. A promising use case for smartwatches is to proactively notify users about recommended items in their vicinity. In this work, we have investigated the usage of smartwatches for displaying proactive recommendations and offering quick feedback options. In order to assess the user acceptance on smartphone, smartwatch and a combination of both devices in a study, we developed a corresponding smartphone and smartwatch application for restaurant recommendations. The results from the study showed that participants preferred notifications on the smartwatch with gesture-based feedback over other options.
Despite the technological advancement of modern hearing aids, many users leave their devices unused due to little perceived benefit. This problem arises from the limitations of the current fitting procedure that rarely takes into account 1) the perceptual differences between users not explained by measurable hearing loss characteristics and 2) the variation in context-specific preferences within individuals. However, the recent emergence of smartphone-connected hearings aids opens the door to a new level of context awareness that can facilitate dynamic adaptation of settings to users' changing needs. In this position paper, we discuss how user auditory intents could be modeled as context collected via mobile devices and suggest what kinds of contextual information are relevant when learning situation-specific intents and the corresponding preferences of hearing impaired users. Finally, we illustrate our ideas with several examples of real-life situations experienced by subjects from our study.
In the domain of human behavior prediction, next-place prediction is an active research field. While prior work has applied sequential and temporal patterns for next-place prediction, no work has yet studied the prediction performance of combining sequential with temporal patterns compared to using them separately. In this paper, we address next-place prediction using the sequential and temporal patterns embedded in human mobility data that has been collected using the GPS sensor of smartphones. We test five next-place prediction methods, including single pattern-based methods and hybrid methods that combine temporal and sequential patterns. Instead of only examining average accuracy as in related work, we additionally evaluate the selected methods using incremental-prediction accuracy on two publicly available datasets (the MDC dataset and the StudentLife dataset). Our results suggest that (1) integrating multiple patterns is not necessarily more effective than using single patterns in average prediction accuracy, (2) most of the tested methods can outperform others for a certain time period (either for the prediction of all places or each place individually), and (3) average prediction accuracies of the top-three candidates using sequential patterns are relatively high (up to 0.77 and 0.91 in the median for both datasets). For real-time applications, we recommend applying multiple methods in parallel and choosing the prediction of the best method according to incremental-prediction accuracy. Lastly, we present an expert tool for visualizing the prediction results.
We present a study addressing the questions of how varying communication styles of a spoken user interface are perceived by users and whether there exist global preferences in the communication styles elaborateness and indirectness. A total of 60 participants had two conversations each with Amazon's Alexa where Alexa used varying wordings for its output. In a post-survey, the participants had to rate statements to subjectively assess each dialogue as well as indicate which dialogue they preferred. The results show that the system's communication style has a direct influence on the user's satisfaction level as well as the user's perception of the dialogue and imply that the preference in the system's communication style is individual for every person. This emphasises the need for adaptive user interfaces.
This paper addresses the design of a model for Question/Answering in an interactive and mobile learning environment. The learner's question can be made through vocal interaction or typed text and the answer is the generation of a personalized learning path. This takes into account the focus and type of the question and some personal features of the learner extracted both from the question and prosodic features, in case of vocal questions. The response is a learning path that preserves the precedence of the prerequisite relations and contains all the relevant concepts for answering the user's question. The main contribution of the paper is to investigate the possibility to exploit educational concept maps in a Q/A interactive learning system.
Recommendation techniques in scientific paper recommender systems (SPRS) have been generally evaluated in an offline setting, without much user involvement. Nonetheless, user relevance of recommended papers is equally important as system relevance. In this paper, we present a scientific paper recommender system (SPRS) prototype which was subject to both offline and user evaluations. The lessons learnt from the evaluation studies are described. In addition, the challenges and open questions for multi-method evaluation in SPRS are presented.
In this paper, we propose to evaluate recommender systems by conducting both offline and user-centric evaluations, while considering multiple quality aspects in realistic settings. This comprehensive evaluation would provide insight on how to improve the algorithms, and how to design better evaluation metrics, particularly for offline settings where it is cheaper to conduct evaluations. We present the preliminary offline evaluation results of several algorithms, using accuracy, novelty, and coverage metrics while considering the impact of dataset density. We propose to complement this offline evaluation with a user-centric evaluation that measures the users' perceived quality of the same algorithms.
The increased utilisation of adaptive visualisation techniques in commercial software requires design principles that provide guidance on how to design user interfaces that change their appearance during runtime. To contribute to the ongoing adaptive user interface (AUI) research, the development of a model is proposed that enables the methodological development of design principles for AUIs. The adoption of the conceptual framework of design-science is proposed for that purpose. Following the framework, the literature of AUI research (to identify observed user behaviour) as well as the literature of human factors (to explain the observed behaviour) have been reviewed. Design principles are built based on this knowledge. This thesis can contribute to an quality improvement of AUI systems.
Machine learning? is a powerful tool in modeling historic user data and converting it into a computational user model. However, it is nearly impossible for a machine learning process to have access to all data pertaining to a user. This leads to potential gaps in the user model produced. Thus, involving the user in overseeing and controlling the modelling process may be considered as a worthwhile goal. This form of user model control is often referred to as scrutability in the personalization research domain. The approach proposed in this research is to combine the benefits of machine learning driven user modelling and scrutable user control in the modelling process. This PhD is at its midpoint and is currently focused on building the model while taking into consideration how user feedback may be incorporated. This feedback is solicited iteratively, enabling the ML process to retrain with the benefit of the user input. The early results of the experimental work show promising prediction accuracy results.
Information visualization is one of the major approach to analyse data. Though there are lot of visualization techniques, designing an adaptive visualization technique for different user and task characteristics is challenging. In this dissertation we are comparing different visualization techniques to find an optimal way for authoring, displaying datasets for two case studies - a crowd sourcing platform for people with different range of abilities and a sensor dashboard for a smart manufacturing set up. We also aspire to develop a user adaptive visualization system. A pilot study found that for numeric dataset, a Bar graph has maximum correct response and Area graph has lowest response time.
Email is an important and widely used communication medium. However, email is increasingly unreliable as people become unlikely to respond to the growing influx of information they receive. Low response rate to email becomes a problem in situations where closing the feedback loop is critical, such as in education, marketing or research. To investigate ways of increasing email response rate, we designed experiments that manipulated the textual elements of the emails. We conducted experiments in a MOOC setting, with email surveys sent out to over 3,000 learners. The emails were sent to elicit responses as to why learners were not engaging with the course. We found that response rates were significantly increased by varying how closely emails were framed as pertaining to a learner's personal situation, such as by changing introductory message, and the format in which links to a survey were presented. Our results yield useful implications to educational and marketing context.
In this paper we introduce the concept of holistic recommendations, namely a set of suggestions generated by exploiting a more comprehensive representation of the user that relies on the personal information coming from different heterogeneous data sources (e.g., social networks, wristbands, smartphones, etc.) and considers the diverse relations and constraints among the data encoded in the profiles. Specifically, in this article we provide the following contributions: i) we outline a conceptual model for providing holistic recommendations built on the ground of such richer user profiles; ii) we present some challenges related to holistic recommendations that can inspire further research in the field.
Historically, UMAP provides the means for usable human-computer interactions; mostly when those take place in dynamic and complex digital environments. However, we are currently witnessing a paradigm shift on how users communicate with each other over the machines e.g., rich contexts, mixed realities, uncertain low information settings. Thus, it seems that traditional practices lack the inherent dynamicity to address adequately these matters and to offer adaptive optimal solutions for enhancing user experience. This paper embraces the existing challenges and suggests an alternative viewpoint by combining concepts from the field of Computational Game Theory. It shifts the focus in the interplay of two (or more) ends while executing a single activity, for identifying the Equilibrium State of Interaction that discloses a win-win situation for all. It consists of suboptimal (best-fit) adaptivity states created dynamically during an interaction hard to predict but able to support a continuous end-to-end engagement.
In this paper we present Myrror, a platform that supports the creation of a unique representation of the user that encodes several facets such as her interests, activities, habits, mood, social connections and so on. Such a representation, that we called holistic user model, is based on the footprints the user spread on social networks and through personal devices. Specifically, our platform acquires personal data coming from several sources, such as Twitter, Facebook, Instagram, Android smartphones and FitBit wristbands, and merges all these information to infer high-level features and populate the facets of the model. Such holistic user models are made available to both the user itself and to third-party services. In the former case, data are shown through a visual interface to improve her self-awareness and her consciousness. In the latter, data are exposed to developers and new personalized services based on these richer user profiles can be created. In both cases, the user has full control over the information she wants to share and unveil.
Modelling physical activity is a recent research topic in the field of user modelling. In our research, we aim to model complex activities such as those required to master the practice of a martial art. In particular, in this paper we present the system MyShikko, aimed to provide a personalized multisensorial support when practicing knee walking, a movement in Aikido martial art that requires keeping the body center aligned when moving the knees on the floor. The motion information for the modelling is collected from a pair of webcams in a stereoscopic configuration and the inertial sensors of a smartphone.
Participating in kindness activities, being generous and showing gratitude, can help people increase their overall happiness levels and improve their levels of wellbeing. This paper describes an installation which will run during the UMAP 2018 Conference. We demonstrate how a gamified digital behavior change intervention can be used to encourage people of different personality types to engage in simple acts of kindness. The system will have implications for future work on personalising persuasive interventions for wellbeing and developing user models through self-assessment and objective behavioural outcomes.
Gamified interfaces have the potential to motivate children's everyday physical activity. An activity model should allow for unobtrusive data input and offer interpretable output for gamified feedback. We define a physical activity user model that uses smartphone sensor data to learn the children's activity level according to classes of MET scores. We conduct a user study using a pre-/post-test design with 61 children to measure the accuracy of the the predicted activity classes compared to accelerometer data. 19 of these children receive gamified feedback on their activity level using a mobile application. We analyze how the effect of the motivational interface compares to influence factors. This work combines user activity modelling with motivational interfaces and offers insights into the limits and chances of applications designed for children.
Current research has shown that a person's personality can be derived from written text on Facebook or Twitter, as well as the amount of information shared on their personal social network sites. So far, there has been no further investigation on whether a person's privacy measures can be extracted from these information sources. We conducted an explorative online user study with 100 participants; the results indicate that privacy concerns can be derived from written text, with a prediction precision similar to personality. At the end of the discussion, we give specific guidelines on the choice of the correct data source for the derivation of the different privacy measures and the possible applications of those.
This paper presents a novel personalized service to support people with Autism Spectrum Disorder (ASD) in their daily movements within urban environments. It aims at helping them in managing anxiety and stress originated by routine breakdowns and unfamiliar situations, improving their autonomy. In particular, we provide an interactive map that is i) personalized, i.e., able to recommend "Safe" Point-of-Interests (SPOIs) according to users' preferences for and aversions to places' sensory features; ii) crowdsourced, i.e., populated with comments and reviews by people with ASD and their caregivers; iii) assistive, i.e., able to support everyday movements (home-work, home-relatives, etc.) by presenting "safe" paths to reach desired places. In this paper we will focus on the first aspect of the map, i.e. SPOIs recommendations. The presented approach is part of a wider project, PIUMA, Personalized Interactive Urban Maps for Autism, which has the aim to develop novel digital solutions for helping people with cognitive problems in their everyday movements.
In this paper we present a deep content-based recommender system (DeepCBRS) that exploits Bidirectional Recurrent Neural Networks (BRNNs) to learn an effective representation of the items to be recommended based on their textual description. Next, such a representation is extended by introducing structured features extracted from the Linked Open Data (LOD) cloud, as the genre of a book, the director of a movie, in order to enrich the available content with new and very descriptive information. In the experimental session we evaluate the effectiveness of our approach in a top-N recommendation scenario: first, we compare the representation based on BRNNs to that obtained through different deep learning techniques. Next, we demonstrate that the exploitation of features gathered from the LOD cloud improves the overall accuracy of our DeepCBRS.
In some scenarios, like music, people often consume items in groups. However, reaching a consensus is difficult, and often compromises need to be made. Such compromises can potentially help users expand their tastes. They can also lead to outright rejection of the recommended items. One way to avoid this is to explain recommendations that are surprising, or even expected to be disliked, by an individual user. This paper presents an approach for generating explanations for groups. We propose algorithms for selecting a sequence of songs for a group to consume. These algorithms consider consensus but have different trade-offs. Next, using these algorithms we generated explanations in a layered evaluation using synthetic data. We studied the influence of these explanations in structured interviews with users (n=16) on user satisfaction.
This paper presents a comprehensive analysis on social media use and engagement by age and gender on Instagram. We define five user age groups (from 10s to 50s) and two user gender groups (males and females), and compare them based on three aspects: activity, image object, and tag. We especially excluded the information that indicates human (e.g., selfies, faces, body) for each aspect in order to examine whether users are still identifiable without that information. Our study results indicate that each user group exhibits unique characteristics and the features from each aspect can be used to develop user classification models (82% for gender and 43% for age classification) without relying on the information that specifically indicates age and gender.
Location Based Social Networks (LBSN) benefit the users by allowing them to share their locations and life moments with their friends. The users can also review the locations they have visited. Classical recommender systems provide users a ranked list of single items. This is not suitable for applications like trip planning,where the recommendations should contain multiple items in an appropriate sequence. The problem of generating such recommendations is challenging due to various critical aspects, which includes user interest, budget constraints and high sparsity in the available data used to solve the problem. In this paper, we propose a graph based approach to recommend a set of personalized travel packages. Each recommended package comprises of a sequence of multiple Point of Interests (POIs). Given the current location and spatio-temporal constraints, our goal is to recommend a package which satisfies the constraints. This approach utilizes the data collected from LBSNs to learn user preferences and also models the location popularity.
The context of collaboration is of great importance. Attempts have been made to objectively define what comprises a successful collaboration. Questions like "When can we say that a collaboration is successful?" or "Is there a way to predict that a collaboration would be successful?" have been asked. In this paper, we look at the output of the collaboration, which are the debugging scores of the pairs, and we consider a collaboration to be successful if it leads to good debugging scores. We choose pair programming because it is an example of a collaboration paradigm. In order to find out what are the potential factors that could possibly predict success in the context of a pair program tracing and debugging task, we performed a dual eye tracking experiment on pairs of novice programmers. We tracked and recorded their fixation sequences and analyzed them using Cross-Recurrence Quantification Analysis (CRQA). Two machine learning algorithms were used, such as Naive Bayes and Logistic Regression. Our findings reveal that CRQA results alone are inadequate to come up with a model with an acceptable performance. Hence, we added the pairs' proficiency level to the model. Between the two models, the Logistic Regression model turned out to be the better model. However, the performance is still not quite unacceptable to predict success so other features are needed to enhance the model.
We have developed PyKinetic, a mobile Python tutor for novices. We present our study on PyKinetic with various activities to target several skills: code tracing, debugging, code understanding and code writing. We compared a version with a fixed sequence of learning activities to an adaptive version, containing the same activities but with personalized problem selection. We had two hypotheses: (H1) the combination of activities is effective for learning, and (H2) the adaptive problem selection is beneficial. The results show that PyKinetic is effective for learning, and the adaptive version provides additional benefits for learners.
In recent years, enterprise group chat collaboration tools such as Slack, IBM's Watson Workspace and Microsoft Teams, have presented unprecedented growth. With all the potential benefits of these tools " productivity increase and improved group communication " come significant challenges. Specifically, users find it hard to focus their attention on content that is relevant to them due to the load of conversational content. This load can be handled by personalized content presentation and summarization mitigated by user profiling. We present an unsupervised approach for implicitly modeling group chat users through a combination of a probabilistic topic model and social analysis. We evaluate our approach by testing it on a task of conversation participation prediction, serving as a proxy for anticipating user interests, and show that by utilizing our approach, a system successfully predicts users participation in conversations. We further analyze the contribution of the various user model components and show them to be significant.
This study explores task persistence in the context of Learning by Teaching. Using features extracted from students' interaction logs, a centroid based clustering algorithm derived two well-separated groups describing two types of students, Cluster 1 which is characterized by the more persistent students and Cluster 0 which is characterized by the less persistent students. The more persistent students demonstrated effective help-seeking behavior, and greater level of task engagement and resourcefulness compared to the less persistent students.
Large, high-resolution displays are highly suitable for creation of digital environments for co-located collaborative task solving. Yet, placing multiple users in a shared environment may increase the risk of interferences, thus causing mental discomfort and decreasing efficiency of the team. To mitigate interferences coordination strategies and techniques were introduced. However, in a mixed-focus collaboration scenarios users switch now and again between loosely and tightly collaboration, therefore different coordination techniques might be required depending on the current collaboration state of team members. For that, systems have to be able to recognize collaboration states as well as transitions between them to ensure a proper adjustment of the coordination strategy. Previous studies on group behavior during collaboration in front of large displays investigated solely collaborative coupling states, not transitions between them though. To address this gap, we conducted a study with 12 participant dyads in front of a tiled display and let them solve two tasks in two different conditions (focus and overview). We looked into group dynamics and categorized transitions by means of changes in proximity, verbal communication, visual attention, visual interface, and gestures. The findings can be valuable for user interface design and development of group behavior models.
Explaining automatically generated recommendations has shown to be an effective means for supporting the user's decision-making process and increasing system transparency. However, present methods mostly provide non-personalized explanations that are presented in an unstructured manner. We propose a framework based on Toulmin's model designed to generate explanations in an argumentative style by presenting supportive as well as critical information about recommended items and their features. Existing research suggests that argumentative explanations cannot be assumed as equally effective for everyone. People rather tend to either apply rational or intuitive decision-making styles that determine which kinds of information are preferably taken into account. In an experimental user study, we investigated the effectiveness of argumentative explanations while considering the moderating effect of these two different cognitive styles. The results indicate that argumentative explanations, as compared to baseline methods, lead to, among others, increased perceived explanation quality, information sufficiency and overall satisfaction with the system. However, this seems only to be true for intuitive thinkers who rely more on explanations in complex decision situations as compared to rational thinkers.
Mobile devices enable users to access information ubiquitously, including in the online learning scenario. This though requires users to multitask and divide their attention between several tasks at once whilst "on-the-go" (e.g. watching a video, walking down the street and keeping track of the traffic at the same time). In order to accommodate learners in this situation, most of today's Massive Open Online Course (MOOC) platforms provide mobile access to their content. Prior works have conducted lab studies to investigate the impact the learning condition (in particular stationary vs. on-the-go) has on mobile MOOC learners. User studies beyond the lab setting though are scarce. We here describe a study in a more realistic setup where 36 participants each participated in two mini-MOOCs while in a stationary and real-life mobile learning situation. We find participants' learning gains slightly lowered in the on-the-go condition (-7%). We also find that on average participants spend 10% more time on video-watching and 23% less time on question-answering in the learning on-the-go compared to the stationary condition.
Group Recommender Systems aim to support the identification of items that best fit individual preferences of group members. However, decision making behavior of group members can be affected by decision biases which can deteriorate group decision quality. In this paper, we analyze the existence of Group Polarization Effects in two different domains and present a way to counteract these effects. Group Polarization is the tendency of a group to make decisions that are more extreme than the average of individual group members' preferences. We analyze Group Polarization in the context of risk analysis and cost estimation. In risk related group decisions, we figured out that if individual group members tend to make cautious decisions, then the group decision will be more cautious. However, in decisions related to cost estimation, the group estimations are lower than the average of group members' estimations (i.e., cautious shift). Furthermore, our results show that individual group members with diverse preferences are not influenced by Group Polarization Effects. The diversity in preferences of individual group members helps to counteract Group Polarization Effects.
Multi-Objective Recommender Systems (MO-RS) consider several objectives to produce useful recommendations. Besides accuracy, other important quality metrics include novelty and diversity of recommended lists of items. Previous research up to this point focused on naive combinations of objectives. In this paper, we present a new and adaptable strategy for prioritizing objectives focused on users' preferences. Our proposed strategy is based on meta-features, i.e., characteristics of the input data that are influential in the final recommendation. We conducted a series of experiments on three real-world datasets, from which we show that: (i) the use of meta-features leads to the improvement of the Pareto solution set in the search process; (ii) the strategy is effective at making choices according to the specificities of the users' preferences; and (iii) our approach outperforms state-of-the-art methods in MO-RS.
Sensor stream data, particularly those collected at the millisecond of granularity, have been notoriously difficult to leverage classifiable signal out of. Adding to the challenge is the limited domain knowledge that exists at these biological sensor levels of interaction that prohibits a comprehensive manual feature engineering approach to classification of those streams. In this paper, we attempt to enhance the assessment capability of a touchscreen based ratio tutoring system by using Recurrent Neural Networks (RNNs) to predict the strategy being demonstrated by students from their 60hz data streams. We hypothesize that the ability of neural networks to learn representations automatically, instead of relying on human feature engineering, may benefit this classification task. Our RNN and baseline models were trained and cross-validated at several levels on historical data which had been human coded with the task strategy believed to be exhibited by the learner. Our RNN approach to this historically difficult high frequency data classification task moderately advances performance above baselines and we discuss what implication this level of assessment performance has on enabling greater adaptive supports in the tutoring system.
The recommendation problem in the hotel industry introduces several interesting and unique challenges leading to the insufficiency of classical approaches. Traveling is not a frequent activity and users tend to have multifaceted behaviors affected by their specific context. While context-aware recommender systems are a promising way to address this problem, the context's dimensions do not contribute equally to the decision-making process and users are not equally sensible to all of the dimensions. In this paper, we propose novel context-aware methods for addressing the hotel recommendation problem, taking into account geography, temporality, textual reviews extracted from social media, and the trip's intent. We present the architecture of the system developed in industry, combining the proposed approaches and addressing each user segment differently. Our experiments show the impact of considering contextual data, external data, and user segmentation on improving the quality of recommendation.
Personalizing Persuasive Technologies (PTs) increase their effectiveness at motivating desired behavioral change. However, most existing efforts towards personalizing PTs and developing personalization models were focused on people from the western countries. In this work, we focused on African audience to investigate how individual's responsiveness to three persuasive strategies (Reward, Social Learning, and Social Comparison) varies by Gender and Age group via a large-scale study of 712 participants. The results of a RM-ANOVA show significant differences in responsiveness to the strategies across the gender and age groups. Females are more responsive to the Reward and Social Learning strategies while males are more responsive to the social comparison strategy. People who are under 25 years are more likely to be persuaded by the Reward and social Learning than participants above 35 years who are more responsive to the Social Comparison strategy. The results will inform PT designers on the appropriate strategy to employ to personalize PTs to individual users based on their Age and Gender.