It is our great pleasure to welcome you to the UMAP 2017 LBR, Demo, and TOR Track, the 25th User Modelling, Adaptation and Personalization, held in Bratislava, Slovakia organized between the July 9-12th, 2017.
This track of UMAP wraps three categories: (i) Demos, which showcase research prototypes and commercially available products of UMAP-based systems, (ii) Late-breaking Results, which contain original and unpublished accounts of innovative research ideas, preliminary results, industry showcases, and system prototypes, addressing both the theory and practice of UMAP and (iii) Theory, Opinion and Reflection (TOR).
TOR is an additional category introduced in this edition of UMAP. Papers in this category critically look at ongoing research topics, reflect on persistent or fleeting trends in the field and offer blue sky future agendas for UMAP research. A novelty related to the TOR category is the presentation format. TOR papers will be presented in the form of an interactive campfire session in a discussion corner during the poster session.
In total, we received 13 LBR, 6 TOR, and 4 Demo submissions out of which 8, 3 and 3 were deemed of high quality by the reviewers. Further 6 LBR papers were accepted from the main conference.
Formative feedback is essential for learning computer programming but is also a challenge to automate because of the many solutions a programming exercise can have. Whilst programming tutoring systems can easily generate automated feedback on how correct a program is, they less often provide some personalised guidance on how to improve or fix the code. In this paper, we present an approach for generating hints using previous student data. Utilising a range of techniques such as filtering, clustering and pattern mining, four different types of data-driven hints are generated: input suggestion, code-based, concept and pre-emptive hints. We evaluated our approach with data from 5529 students using the Grok Learning platform for teaching programming in Python. The results show that we can generate various types of hints for over 90% of students with data from only 10 students, and hence, reduce the cold-start problem.
The research on group recommender systems is often oversimplifying the problem of generating group recommendations, as it is usually only considering the explicit preferences of the group members and, in some cases, enriching these preferences with additional information about the individual members. In this way, an essential aspect is frequently completely neglected: the characterization of the group as an entity with a specific composition and with group-related dynamics. The goal of this paper is multifaceted, firstly, to address the limitations of state-of-the-art approaches, secondly, to describe the problem of group recommendations in a more comprehensive fashion, thirdly, to summarize the results of our previously conducted analyses as a supporting evidence of a need for richer group models, and finally, to discuss an alternative and rather novel approach to group recommendations in the tourism domain. To this end, the results of the group decision-making study with 200 participants in 55 groups are summarized and related to the seven travel factors of the picture-based recommendation system.
The large volume of data available in location-based social networks (LBSNs) enables Point-of-Interest (POI) recommendation services. On another hand, the heterogeneous information (e.g., user check-in records, geographical features of POIs, social network and user reviews) imposes great challenges on effective POI recommendation. In this paper, we focus on leveraging such rich information in an integrated manner to improve POI recommendation performance. We exploit not only the geographical and social information, but also aspects extracted from user reviews to better model users' preferences. More specifically, to fully utilize various types of information, we construct a novel heterogeneous graph, Aspect-aware Geo-Social influence Graph (AGSG), by fusing various relations among the three types of nodes, i.e., users, POIs and aspects. The personalized POI recommendation task is then transformed as a graph node ranking problem in AGSG. We design a graph-based recommendation algorithm based on both personalized PageRank (PPR) and meta paths, to fully exploit the heterogeneous graph structure as well as to capture the semantic relations among the various nodes. Experiments on three real-world datasets show that our proposed approach outperforms the state-of-art methods.
Recommender systems have become important tools to support users in identifying relevant content in an overloaded information space. To ease the development of recommender systems, a number of recommender frameworks have been proposed that serve a wide range of application domains. Our TagRec framework is one of the few examples of an open-source framework tailored towards developing and evaluating tag-based recommender systems. In this paper, we present the current, updated state of TagRec, and we summarize and reflect on four use cases that have been implemented with TagRec: (i) tag recommendations, (ii) resource recommendations, (iii) recommendation evaluation, and (iv) hashtag recommendations. To date, TagRec served the development and/or evaluation process of tag-based recommender systems in two large scale European research projects, which have been described in 17 research papers. Thus, we believe that this work is of interest for both researchers and practitioners of tag-based recommender systems.
In the digital era, personalisation systems are the typical way to deal with the massive amount of information on the Web. ese systems decide in our place what we like, possibly hiding us away from a complete world of potentially interesting content. ese systems do not challenge us to open our horizons of interest, trap- ping us more and more in our lter bubble. Introducing diversity and serendipity in the recommendation results has been widely recognised as the solution to this issue in the information retrieval eld. However, serendipity cannot be addressed and measured with traditional accuracy metrics, because it introduces much more complexity in terms of subjectivity and personality. Inspired by the curiosity theory of Berlyne, further developed by Silvia, we introduce in user pro les a so-called coping potential estimation as a measure of the users' ability to cope with new items (e.g., ability to appreciate serendipitous recommendations). Our assumption is that curiosity leads to serendipity and high coping potential users accept more serendipitous results, and thus we need to model it in the recommendation algorithm. We performed an online ex- periment where we asked users a number of questions about TV programmes recommendations. Our results show that users with a high coping potential are more inclined to accept serendipitous recommendations than their counterparts.
Recommender systems generate recommendations by analysing which items the user consumes or likes. Moreover, in many scenarios, e.g., when a user is visiting an exhibition or a city, users are faced with a sequence of decisions, and the recommender should therefore suggest, at each decision step, a set of viable recommendations (attractions). In these scenarios the order and the context of the past user choices is a valuable source of data, and the recommender has to effectively exploit this information for understanding the user preferences in order to recommend compelling items.
For addressing these scenarios, this paper proposes a novel preference learning model that takes into account the sequential nature of item consumption. The model is based on Inverse Reinforcement Learning, which enables to exploit observations of users' behaviours, when they are making decisions and taking actions, i.e., choosing the items to consume. The results of a proof of concept experiment show that the proposed model can effectively capture the user preferences, the rationale of users decision making process when consuming items in a sequential manner, and can replicate the observed user behaviours.
In this paper, we explore the question "would people be willing to share their personal data in exchange for highly-personalized online ads?" through a Wizard-of-Oz deception study. Our volunteers were exposed via a web browser to three different highly- personalized ads, designed by people who knew them well. They were made believe that the ads had been generated automatically by an Artificial Intelligence engine on the basis of their browsing & location history and/or personal traits. The participants' reactions were surprisingly favorable: in more than 50% of the cases, the ads triggered spontaneous positive emotional reactions; almost 90% of participants would share at least two of the three data sources with advertisers; and about 50% would share all data sources. Our results provide evidence that highly-personalized ads may offset the concerns that people have about sharing their personal data. Thus further efforts in building increasingly personalized online ads would represent a worthwhile endeavour.
Research has shown that social influence can be used to effect behavior change. However, research on the role culture plays in the effect of age and gender on social influence in persuasive technology is scarce. To address this, we investigate the effect of age and gender on the susceptibility of individuals to Competition, Reward, Social Comparison and Social Learning in individualist and collectivist cultures, using a sample of 360 participants from North America, Africa and Asia. Our results reveal that there are more significant differences between males and females and between younger and older people in collectivist cultures than individualist cultures. In individualist culture, we found that males and females differ with respect to Competition only, with males being more susceptible. However, in collectivist culture, we found males differ from females with respect to Reward and Competition, with males being more susceptible, while younger people differ from older people with respect to Competition, Social Comparison and Social Learning, with younger people be more susceptible. Our findings provide designers of gamified persuasive applications with empirical insights for tailoring to the different cultures based on age and gender
Research has shown that the perceived credibility of a website is critical to its success. However, little is known about how individual differences influence this important factor of web design. In this paper, we investigate how personality traits affect the perceived credibility of a website in the mobile domain. Using a sample of 323 participants, we developed a model showing how the Big Five personality traits influence the perceived credibility of a website through its perceived aesthetics and perceived usability. Our model reveals that Agreeableness is the strongest predictor of aesthetics and/or usability, followed by Conscientiousness. This suggests that the more agreeable and/or the more conscientious users are easily more satisfied aesthetically and usability-wise by a mobile websites than the less agreeable and/or the less conscientious users respectively. Consequently, designers of mobile sites may have to do more in user interface design in order to attract the less agreeable and/or the less conscientious users to their sites based on its hedonic (aesthetics-inspired) and utilitarian (usability-inspired) appeal.
In this paper, we investigate differences between recipes uploaded by users and recipes bookmarked by users. The results indicate that uploaded recipes outperform bookmarked recipes in terms of healthiness. Further, health scores and nutritional values of these recipes are highly related to the stated cooking interests: for example, Southern Food lovers eat not as healthy as those who prefer the Mediterranean or Middle-Eastern cuisine. A disturbing finding is that interest in the category `Kids' is associated with bad values for all nutritional measures. We also found some interactions between hobbies such as biking, hunting or knitting and nutritional values. These insights pave way to the design of health-aware recipe recommender systems that take a user's food preferences into account; in addition, taking a user's lifestyle and hobbies into account would provide valuable input to persuasive systems.
Increasing diversity in the output of a recommender system is an active research question for solving a long-tail issue. Most of the current approaches have focused on ranked list optimization to improve recommendation diversity. However, little is known about the effect that a visual interface can have on this issue. This paper shows that a multidimensional visualization promotes diversity of social exploration in the context of an academic conference. Our study shows a significant difference in the exploration pattern between ranked list and visual interfaces. The results show that a visual interface can help the user explore a a more diverse set of recommended items.
The intent-aware diversification framework considers a set of aspects associated with items to be recommended. A baseline recommendation is greedily re-ranked using an objective that promotes diversity across the aspects. In this paper the framework is analysed and a new intent-aware objective is derived that considers the minimum variance criterion, connecting the framework directly to portfolio diversification from finance. We derive an aspect model that supports the goal of minimum variance and that is faithful to the underlying baseline algorithm. We evaluate diversification capabilities of the proposed method on the MovieLens dataset.
Many studies have sought to understand the behavior of music listeners to design an improved music listening experience. This is especially important in music recommendation systems in that listening behavior can directly relate to the purpose of the system. For example, a listener who likes to discover new music will be more satisfied with a list of suggestions that present different types of music, while others prefer to listen to their same old music and artists. Previous research has focused on performing user research to explicitly extract information about listening behavior but few studies have attempted a data-driven approach to suggest listener personas or groups. In this study, we applied two clustering methods to user playrate distribution data to see if meaningful user clusters appear, and performed analysis on the results by comparing the patterns of the result clusters with the major characteristics of listener groups derived from previous user researches. Our experiments show that two large clusters and two small clusters are formed, with each cluster representing an intuitive difference in terms of listening behavior.
Massive Open Online Courses (MOOCs) play an ever more central role in open education. However, in contrast to traditional classroom settings, many aspects of learners' behaviour in MOOCs are not well researched. In this work, we focus on modelling learner behaviour in the context of continuous assessments with completion certificates, the most common assessment setup in MOOCs today. Here, learners can obtain a completion certificate once they obtain a required minimal score (typically somewhere between 50-70%) in tests distributed throughout the duration of a MOOC. In this setting, the course material or tests provided after "passing" do not contribute to earning the certificate (which is ungraded), thus potentially affecting learners' behaviour. Therefore, we explore how ``passing'' impacts MOOC learners: do learners alter their behaviour after this point? And if so how? While in traditional classroom-based learning the role of assessment and its influence on learning behaviour has been well-established, we are among the first to provide answers to these questions in the context of MOOCs.
Psychomotor learning is crucial for many kind of tasks that involve the acquisition of motor skills, such as practicing martial arts. In the past two decades, diverse technological solutions have been developed to support the learning of the corresponding motor skills. However, the UMAP (User Modeling, Adaptation and Personalization) community has not taken part in that research efforts and thus, resulting systems do not adapt and personalize the system response to their users' needs. This paper discusses the main features of existing systems to learn martial arts and identifies research opportunities that are to be included in the future agenda for UMAP research.
In order to use and model nutritional knowledge in a food recommender system, uncertainties regarding the users nutritional state and thus the personal health value of food items, as well as conflicting nutritional theories need to be quantified, qualified and subsumed into falsifiable models. In this paper, we reflect on different error sources with respect to nutrition and consider how such issues can be tackled in future systems. We discuss the integration of general nutritional theories into information systems as well as user specific nutritional measures and different approaches to evaluating the utility of a given nutritional model.
The notion of intelligent systems that "care" is at the center of research in areas such as Intelligent Tutoring Systems and Adaptive Systems. This paper elaborates on the notion of caring assessment systems, and presents work towards achieving this vision that has potential for improving students' assessment experiences.
This demo paper describes the semantic query interpretation model adopted in the OnToMap Participatory GIS and presents its benefits to information retrieval and personalized information presentation.
In partner dance classes, teachers typically manage several students at the same time. For that reason, the amount of feedback provided in class is quite limited and students do not have many resources to get other feedback. In this demo paper, we present Forró Trainer, a tool that allows students to practice dance exercises by themselves, receiving automatically generated feedback about their performance. The system runs on a smartphone app and focuses on a fundamental aspect of dancing: learning to follow the rhythm of the music. The app detects the student's movements, using the mobile's accelerometer, extracts aspects of the rhythm and provides feedback. We present a description of the tool, the mistakes it detects, the automated feedback and the benefits that it may provide for dance students.
This demo paper presents a Web application for recommending travel regions for independent travelers. Users can specify preferences such as budget and preferred activities and receive suggested trips consisting of multiple regions. We explain the main ideas behind the data model and algorithm of our solution, and give an overview on the implementation
My doctoral research will investigate adapting group formation in computer supported collaborative learning (CSCL) based on learners' characteristics. As Group based learning leverages on interaction for effective cognition, this project aims to investigate the effect of individual behavioural characteristics on interaction within a group. Based on our findings, we will develop and evaluate a model for adapting group formation for effective interaction in CSCL.
News recommendation poses several specific challenges compared to other domains, such as freshness and serendipity. The proposed research will develop new methods and techniques to address some of such challenges. With the aim of handling the users' changing interests and the fast evolution overtime of news, my solution will be proposed in the social network domain, exploiting an adaptive focused crawling algorithm. Moreover, it will consider a given user's attitude towards her interests, with the purpose of recommending articles in line with her beliefs. An experimental evaluation is currently being implemented to assess the effectiveness of my approach, also in comparison with state-of-the-art techniques.
Early studies about user modeling already noted that user models need to continuously adapt, as new information about the user becomes available, keeping the user model updated. Although temporal aspects are inherent characteristics of user modeling, so far, they were not specifically dealt with. Therefore, the proposed research will suggest an abstract, generic framework to deal with temporal aspects of user modeling while providing guidelines based on domains characteristics for handling temporal aspects in User Modelling.
Recommender systems are used to suggest users products that they would not be able to find by themselves. State of the art algorithms assume that items have static features, however this assumption does not always correspond to reality. There are challenging and still unexplored domains, where not only users but also items have properties that evolve continuously over time.
In this research we aim to overcome these limitations by suggesting to model evolution of users and items as a reinforcement learning problem. As use case we will refer to the recommendation problem applied to the financial domain, where items' (contracts) features evolve continuously according to "market laws".
Nowadays we assist to a significant innovation of the teaching practises due to the crisis of the classical teaching approach, the availability of low cost mobile technology and the easy access to global knowledge and information. Learning Design systems represent valuable tools to support teachers in the delicate task of organizing the teaching-learning activities in active student-centered approaches. There are many active projects in this field, but the available tools do not always fulfill the expectations. Furthermore, there is a rapid growth of Web 2.0 apps to create digital artefacts with a strong potential impact in learning activities, but current LD platforms don't guide teachers and students in choosing best apps to carry on a specific task. This paper provides an overview of the state of the art LD tools and developing perspective in this area.
Intelligent Content was proposed by Ann Rockley as content that is structurally rich and semantically aware and therefore automatically discoverable, reusable, reconfigurable, and adaptable. Currently, the majority of approaches to achieving Intelligent Content are service-side, where cloud services discover and transform content before delivering it to a client or user. An alternative approach to achieving Intelligent Content, embeds the intelligence within the content itself, imbuing the content with the ability to call services and enact discovery, reusability, reconfiguration and adaption on the client-side.Thus, the Intelligent Content can be proactive in calling cloud services and perform contextually relevant transformations or behaviours.
This work explores this client-side approach to Intelligent Content and aims to examine the content-service interactions in such a system. The research follows a case-study based approach, examining the development of a client-side user model utilised to personalise and cache content on the client-side. Specifically, the research examines a propensity model, monitoring implicit user actions such as mouse movement and scrolling to create content that knows the propensity of a user to click on various page elements. Evaluation of the proposed architecture will examine the impact of content-service interaction frequency on system accuracy and response time.
We are happy to present the 8 workshops and 2 tutorials selected for the 25th ACM Conference on User Modeling, Adaptation, and Personalization (UMAP 2017) from the proposals we received in response to the corresponding calls for workshops and tutorials. The workshops provide a venue to discuss and explore emerging areas of User Modelling and Adaptive Hypermedia research with a group of like-minded researchers and practitioners from Industry and Academia. The discussions are focused on systems that adapt to individual users, to groups of users, and that collect, represent, and model user information. The workshops held in conjunction with UMAP 2017 cover a wide range of topics.
Welcome to the 4th International Workshop on Educational Recommender Systems (EdRecSys) held in conjunction with the 25th ACM Conference on User Modeling, Adaptation, and Personalization (UMAP 2017). Recommender systems have become increasingly popular in recent years, helping us to make decisions about what products to buy, what movies to watch, what books to read or who to date. While these systems have shown their effectiveness in e-commerce, music and social networks, the field of education is an emerging and very promising application area. The educational environment is no longer limited to face-to-face classes; it includes online learning and activities using Technology Enhanced Learning (TEL), Learning Management Systems (LMS) and Massive Open Online Courses (MOOC), all of which can benefit from the application of recommender systems to alleviate information overload and improve personalisation, to better meet the needs of the individual student. For example, high school and university students can be provided with recommendations about: (i) suitable degrees and courses, based on their background, preferences and prior experience; (ii) project and thesis topics, and appropriate supervisors; (iii) internships and jobs; (iv) other students to work with (to do group work or peer learning); (v) suitable learning resources based on their knowledge, skills and learning style, e.g. books, tutorials, activities, etc.
This workshop has brought together researchers and practitioners from the areas of user modeling and personalisation, recommender systems, education, data mining, learning analytics, intelligent tutoring systems and other related disciplines, to explore the use of recommender systems in education, share their experience and discuss the challenges and open research topics in the design and deployment of effective solutions.
Lack of social relationship has been shown to be an important contribution factor for attrition in Massive Open Online Courses (MOOCs). Helping students to connect with other students is therefore a promising solution to alleviate this phenomenon. Following up on our previous research showing that embedding a peer recommender in a MOOC had a positive impact on students' engagement in the MOOC, we compare in this paper the impact of three different peer recommenders: one based on socio-demographic criteria, one based on current progress made in the MOOC, and the last one providing random recommendations. We report our results and analysis (N = 2025 students), suggesting that the socio-demographic-based recommender had a slightly better impact than the random one.
Authoring an adaptive educational system is a complex process that involves allocating a large range of educational content within a fixed sequence of units. In this paper, we describe Content Wizard, a concept-based recommender system for recommending learning materials that meet the instructor's pedagogical goals during the creation of an online programming course. Here, the instructors are asked to provide a set of code examples that jointly reflect the learning goals that are associated with each course unit. The Wizard is built on top of our course-authoring tool, and it helps to decrease the time instructors spend on the task and to maintain the coherence of the sequential structure of the course. It also provides instructors with additional information to identify content that might be not appropriate for the unit they are creating. We conducted an off-line study with data collected from an introductory Java course previously taught at the University of Pittsburgh in order to evaluate both the practicality and effectiveness of the system. We found that the proposed recommendation's performance is relatively close to the teacher's expectation in creating a computer-based adaptive course.
During the phases of course construction, in Learning Management Systems, a teacher can be valuably helped by system's recommendations about learning objects to include in the course. A usual protocol is in that the teacher performs a query, looking for suitable learning material, and the system proposes a list of learning objects, with information shown for each one; then the teacher is supposed to make her choice, basing on the displayed information. Here we present a Recommender System for Learning Objects retrieved from Learning Objects Repositories, that is based on a ``social teacher model", based on the similarities with the teacher in the system, and the potential model evolutions over time. The proposed system is available as a Moodle plug-in.
In the paper we show the details of the information decorating the learning objects retrieved after a query, the definition of the teacher model, and the similarity measure underlying the recommendation strategy.
This paper presents a data driven approach to autonomous course-competency requirement and student-competency level discovery starting from the grades obtained by a sufficiently large set of students. The approach relies on collaborative filtering techniques, more precisely matrix decomposition, to derive the hidden competency requirements and levels that together should be responsible for observed grades. The discovered hidden features are translated into human understandable competencies by matching the computed values to expert input. The approach also allows for grade prediction for so far unobserved student course combinations, allowing for personalized study planning and student guidance. The technique is demonstrated on data from a "Data Science and Knowledge Engineering" Bachelor study, Maastricht University.
Nowadays using E-learning platforms such as Intelligent Tutoring Systems (ITS) that support users to learn subjects are quite common. Despite the availability and the advantages of these systems, they ignore the learners' time limitation for learning a subject. In this paper we propose RUTICO, that recommends successful learning paths with respect to a learner's knowledge background and under a time constraint. RUTICO, which is an example of Long Term goal Recommender Systems (LTRS), after locating a learner in the course graph, it utilizes a Depth-first search (DFS) algorithm to find all possible paths for a learner given a time restriction. RUTICO also estimates learning time and score for the paths and finally, it recommends a path with the maximum score that satisfies the learner time restriction. In order to evaluate the ability of RUTICO in estimating time and score for paths, we used the Mean Absolute Error and Error. Our results show that we are able to generate a learning path that maximizes a learner's score under a time restriction.
The increasing variety of programming languages available to computer programmers has led to the discussion of what language(s) should be learned. A key point in the choice of a programming language is the availability of support from experienced programmers. In this paper, we explore the use of graph theory in recommending programming languages to novice and expert programmers in a question and answer collaborative learning environment, Stack Overflow. Using social network analysis techniques, we investigate the relationship between experts (using an expertise graph) in different programming languages to identify what languages can be recommended to novice and experienced programmers. In addition, we explore the use of the expertise graph in inferring the importance of a programming language to the community. Our results suggest that programming languages can be recommended within organizational borders and programming domains. In addition, a high number of experts in a programming language does not always mean that the language is popular. Furthermore, disconnected nodes in the expertise graph suggest that experts in some programming languages are primarily on Stack Overflow to support that language only and do not contribute to questions or answers in other languages. Finally, developers are comfortable with mastering a single, general purpose language. The results of our study can help educators and stakeholders in computer education to understand what programming languages can be suggested to students and what languages can be taught and learned together.
Individualized and personalized learning has taken on different forms in the context of digital learning environments. In intelligent tutoring systems, individualization is focused on estimation of the cognitive mastery of the student and the speed at which the student progresses through the material is conditioned on her individual rate of mastery. In prior work, a recommendation framework based on learner behaviors, rather than learner's cognitive abilities, was proposed and developed. This framework trained a behavior model on millions of previous student actions in order to estimate how a future learner might behave. This behavior model can incorporate the amount of time spent on each course page, such that the model can take into account a learner's previous behaviors and provide a specific course page recommendation to where the learner may want to go next where they can be expected to spend a significant amount of time on. We stipulate that this approach touches on factors more aligned with personalization, since the prediction of behavior is an aggregation of the student's cognitive abilities, affective state, and preferences. This model was applied to a hand-picked pair of MOOC offerings where model results were expected to be favorable. In this paper, we investigate the suitability of this behavioral prediction approach by applying it to an expanded set of 13 UC Berkeley MOOCs run on the edX platform. Preliminary results from applying the time-augmented Recurrent Neural Network (RNN) based behavior model approach are presented and compared to baseline models. These findings contribute to the discussion of when and in what context this form of collaborative based personalized recommendation is appropriate in MOOCs.
It is our great pleasure to welcome you to the UMAP 2017 EvalUMAP workshop. The purpose of this workshop series is to establish comparative evaluation tasks and a suitable framework to support researchers in the user modelling and personalisation research field in comparing their research to that of others. In 2016 we discussed the key challenges in performing such comparative evaluations. This year we focus on the challenges of identifying appropriate datasets and methods. In particular, the planned outcomes of the workshop this year are as follows: (1) Understand the challenges and requirements related to the design of a shared task approach in User Modeling, Adaptation and Personalization space, (2) identify suitable and publicly accessible datasets that overcome the previous identified challenges and requirements and (3) using the identified datasets start designing shared evaluation tasks that will be performed throughout 2017 and be presented at UMAP 2018.
Evaluation of personalized systems is a complicated endeavor. First, evaluation goals, methods and criteria are manifold and have to be carefully selected to fit the actual application scenario and the scope of the evaluated system. Second, it is considerably harder to locate the source of problems, compared to non-adaptive systems where problems most often reside on the UI level. Thus, in the past, a layered evaluation approach for personalized systems has been proposed that distinguishes between five layers that can theoretically all be the source of problems (e.g., collection of input data or adaptation decision). This paper outlines a use case related to personalized interaction comprising i) modeling a user's interaction abilities, ii) recommending interaction methods and devices that fit the user's individual needs, and iii) personalized system behavior and reaction to user input. Next, the paper describes experiences with an evaluation process using the layered evaluation framework.
Open-source mobile notification datasets are a rarity in the research community. Due to the sensitive nature of mobile notifications it is difficult to find a dataset which captures their features in such a way that their inherently personal information is kept private. For this reason, the majority of research in the domain of Notification Management requires ad-hoc software to be developed for capturing the data necessary to test hypotheses, train algorithms and evaluate proposed systems. As an alternative, this paper discusses the process, advantages and limitations with harnessing a large-scale mobile usage dataset for deriving a synthetic mobile notification dataset used in testing and improving an intelligent Notification Management System (NMS).
Test collections for offline evaluation remain crucial for information retrieval research and industrial practice, yet reusability of test collections is under threat by different factors such as dynamic nature of data collections and new trends in building retrieval systems. Specifically, building reusable test collections that last over years is a very challenging problem as retrieval approaches change considerably per year based on new trends among Information Retrieval researchers. We experiment with a novel temporal reusability test to evaluate reusability of test collections over a year based on leaving mutual topics in experiment, in which we borrow some judged topics from previous years and include them in the new set of topics to be used in the current year. In fact, we experiment whether a new set of retrieval systems can be evaluated and comparatively ranked based on an old test collection. Our experiments is done based on two sets of runs from Text REtrieval Conference (TREC) 2015 and 2016 Contextual Suggestion Track, which is a personalized venue recommendation task. Our experiments show that the TREC 2015 test collection is not temporally reusable. The test collection should be used with extreme care based on early precision metrics and slightly less care based on NDCG, bpref and MAP metrics. Our approach offers a very precise experiment to test temporal reusability of test collections over a year, and it is very effective to be used in tracks running a setup similar to their previous years.
In this paper we present a simple, yet powerful approach to generating labeled datasets of Twitter1 users. Our focus falls on sensitive personal details, shared as background information in tweets. Such tweets avoid the focus of user's attention and also tend to resist the vast amounts of humor, wishes or hypothetical thinking typical for tweets.
Our approach combines selecting search queries, followed up by a semi-supervised filtering of indicative messages. We create datasets in several unrelated domains and prove that all sorts of target groups can be built with minimal manual annotator effort.
The generated datasets include separate groups of users with specific characteristics: pet ownership, blood pressure, diabetes and psychotropic medicine usage, for which to our knowledge manually labeled data was previously not available.
Our search-based approach is also used to generate a cross-domain corpus, matching Twitter users with their Yelp2 profiles.
Evaluation of user modeling techniques is often based on the predictive accuracy of models. The quantification of predictive accuracy is done using performance metrics. We show that the choice of a performance metric is important and that even details of metric computation matter. We analyze in detail two commonly used metrics (AUC, RMSE) in the context of student modeling. We discuss different approaches to their computation (global, averaging across skill, averaging across students) and show that these methods have different properties. An analysis of recent research papers shows that the reported descriptions of metric computation are often insufficient. To make research conclusions valid and reproducible, researchers need to pay more attention to the choice of performance metrics and they need to describe more explicitly details of their computation.
This paper describes and proposes a community evaluation task that is designed for evaluating learning systems that can automatically identify different types of problems, that students may encounter with their online courses. As a basis, the learning systems would use logs from an artificial learning environment to analyse the student interactions and behaviour with the online course. The learning systems will also use specific domain models to ensure that the course requirements such as task deadlines and learning content conditions (e.g., pre-requisites) are addressed. As a result, the outputs (identified student problems) can be used by a) the learning systems to provide personalised feedback and direction to students to overcome a problem b) notify an instructor for a more professional support and response c) inform a learning designer for potential problems on the design of the course.
It is our great pleasure to welcome you to the 2nd International workshop on Human Aspects in Adaptive and Personalized Interactive Environments (HAAPIE 2017). HAAPIE 2017 (http://haapie.cs.ucy.ac.cy) is a full-day workshop held on 09 July 2017 in conjunction with the 25th ACM Conference on User Modeling, Adaptation and Personalization (UMAP 2017), 09-12 July 2017 in Bratislava, Slovakia. 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", 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 for sharing ideas and experiences, lessons learned, approaches and results that could substantially contribute to the broader UMAP community.
This year we received 14 submissions from all around the world covering a broad range of topics on the workshop's research themes areas. Each paper has been reviewed by up to 3 members of the IPC with expertise in the respective area to ensure the necessary relevance, quality and novelty.
Research has shown that persuasive technologies are more effective when they are personalized. Persuasive strategies work differently for various people; hence a one size fits all approach may not bring about the desired change in behavior or attitude. This paper contributes to personalization in question and answer (Q&A) social networks by exploring the possibility of personalizing social influence strategies based on the computer programming skill level and the highest level of education of users. In particular, this paper explores the susceptibility of users in Stack Overflow, a Q&A social network, to social support influence strategies for novice and expert computer programmers. In addition, we explore if first degree holders respond to the social support influence strategies the same way graduate degree holders do. Using a sample size of 282 Stack Overflow users, we constructed four models using Partial Least Squares Structural Equation Modelling (PLS-SEM) and carried out multi-group analysis between these models. The results of our analysis show that social facilitation significantly influences cooperation for novice programmers, but not for expert programmers. In addition, social learning does not significantly influence the persuasiveness of the system for expert programmers compared to users who are novice in computer programming. For the users grouped according to their highest level of education, social learning influenced cooperation among the graduate degree holders and competition influenced the graduate degree holders to continue using the system. The result of this study can provide useful guidelines to social network developers that can be used in implementing personalized influence strategies in Q&A social communities.
Personalization has been discussed in a number of domains such as learning, search or information retrieval. In the area of human-computer interaction, personalization also plays a prominent role. For the variety of users, especially for those with impairments, interaction abilities vary drastically and for many it is not possible to use common interaction devices like a mouse or the touch screen of a smart phone (at least not with their predefined configuration). This paper describes a personalized interaction approach based on i) modeling users' interaction abilities, ii) automated selection and configuration of interaction devices for the individual user, and iii) adaptive behavior of selected applications. It is not exclusively focused on the requirements of people with impairments but frequently takes up this target group as it might particularly benefit from interaction personalization concepts.
Group Recommendation Systems (GRS) are personalization systems that provide recommendations to groups of people considering the initial preferences of each group's member, with the aim to maximize the satisfaction of the whole group. Since recent psychological studies evidence that people's satisfaction is influenced by the satisfaction of other people with whom they perform an activity, it is important to consider human aspects and social characteristics that affect the changes in individual's satisfactions in the recommendations generation process. In this work, we start an experimental analysis on how ties' strength and possible conflicts in a relationship can influence the individual's satisfactions, with the aim to derive a model that can be used to adapt individual utilities to the "Group Context" before aggregating them into the group's ones. Our hypothesis is that there is a direct correlation between tie strength and positive shifting, but the presence of conflict, instead, can lead to a negative influence, causing a drifting further apart between people's satisfactions. Results confirm these hypotheses, but also suggest that these two factors are not enough to define a general model and that other factors must be considered.
In the past decades, the process of urbanization has shaped general socio-economic aspects of cities with different population sizes. Among them, food consumption is a good indicator to reflect the quality of life. In this paper, we study the impact of city size on food preferences, as shown by users of a large German food sharing community. We quantitatively and qualitatively analyze differences in dietary choices made by users who indicate to live in cities of different sizes, from metropolises and big cities to medium and small towns. Further, we demonstrate that the city size of the creators of online recipes can be predicted with a good accuracy of 86%, using predictors based on recipe authors' profiles, recipe popularity, season, and recipe complexity and contents. The findings indicate that city size is a useful feature to take into account in various other domains.
This paper presents principles of designing Educational Assessment Technology (EdAT) that is culturally-appropriate. These principles are based on a study on international large-scale assessments (ILSA) and the relations of their findings with culture. In order to achieve this, we correlate ILSA data with Cultural Dimension data for 81 countries and examine the implications for designing student-centred systems based on cultural dimensions. Cultural dimensions such as long-term orientation are good predictors for achievement and can guide design decisions around adaptation.
Learning through video watching has been popular through the education community and is considered as a common choice especially for self-directed informal learning. However, the learner in this situation acts only as a passive consumer and does not receives any feedback for improving his/her performance, an element important in any educational context. Studies in music psychology reveal that gender, perceptual, and cognitive, differences, along with the level of music education of the individual, should be considered when support is generated to a person who is watching a music video for educational purposes. In this sense, individual differences should be exploited when designing an adaptive learning support aiming to suit the individual in music learning. In this line, this paper presents an exploratory study into interaction data of music experts and amateurs when they were actively watching a music video. Linguistic analysis is also employed for taking an insight into the written comments provided by participants at several timepoints in the music videos. Results reveal significant differences between genders in their interaction behavior but also in their perception processing of the music videos, reflected in their comments. Suggestions are provided based on the results on how these can be utilized for the design of personalized support in informal music education.
Although there is abundant evidence that individual differences such as cognitive abilities impact visualization effectiveness, this influence has mostly been shown for fictional tasks/scenarios. This paper extends previous findings by investigating the impact of individual differences on user experience with a real-world information visualization tool designed to support preferential choices in public engagement. We show that several cognitive abilities do have an influence on user experience in this task, and show that this influence can be explained by eye tracking. We discuss how these results are promising towards the design of visualizations for preferential choice in public engagement that can adapt to the user's needs, abilities and expertise.
We introduce a new probabilistic working memory (WM) model that we intend to use to automatically personalize user interfaces with respect to Alzheimer patients' declining WM capacity. WM is the part of the human memory responsible for the conscious short-term storing and manipulation of information. It is known to be extremely limited and to be one of the strongest factors that impact individual differences in cognitive abilities. In particular, individuals suffering from Alzheimer's disease have significantly impaired WM capacities that worsen as the disease progresses. As a use case for our model, we describe a system that is designed to help patients with Alzheimer's disease choose the music track they would like to listen to from a given playlist. We discuss how our WM model could be used to adapt this system to each patient's disease progression in time and the consequent deterioration of her WM capacity.
Public interactive displays (PIDs) are becoming more pervasive in urban environments as a means to engage passers-by and to provide interactive features such as wayfinding. However, one of the problems with current PIDs is that they are typically designed around an average specification, potentially excluding a large range of users that for instance might not be able to reach interactive elements. To address this challenge, we propose a number of design concepts for adjusting PIDs to users of different heights. We present a preliminary evaluation of our concepts through a cognitive walk-through study with 10 design experts using a custom-developed experience prototype featuring four height-aware modes. Based on qualitative feedback and observations we discuss design suggestions for future work.
In a rapidly growing mobile device market, a new and unique type of IT product, the smartwatch, has started gaining the users' attention. After several years of technological development, it has finally become a viable device that extends the functions of a smartphone to a more intimate level. In the industry of smart devices, particularly in wearable devices, smartwatches are widely considered as the next big thing which is going to have a significant impact on our daily lives. However, smartwatches still have a limited use in public transport information systems. In this paper, we present promising use cases for smartwatches in public transport which were extracted through a survey questionnaire we conducted. We designed and developed a prototype to realize one of the most promising use cases: real-time public transport navigation. We evaluated our prototype in a comprehensive user study. The comparison of the smartwatch-based navigation with a pure smartphone-based solution shows that the smartwatch outperforms the smartphone in all user experience metrics.
In this paper we describe a preliminary investigation in using pupil dilation measurements to understand user visualization processing, with the long-term goal of building user-adaptive visualizations that can tailor the presentation of complex visual information to specific user needs and states. In particular, we look at how a selection of pupil dilation measurements are affected by adding several highlighting interventions designed to aid visualization processing to a bar graph.
This paper investigates the influence of learner personality and learning styles on the selection of different styles of learning materials. We considered the big five personality traits (focusing in particular on Extroversion and Openness to Experience) and Felder and Soloman's Index of Learning Styles instrument (ILS). We found no real impact of learning styles, except for a small effect for the visual/verbal style. We also did not find an impact of personality on the selection of different styles of learning materials.
Systems for Community Question Answering (CQA) are well-known on the open web (e.g. Stack Overflow or Quora). They have been recently adopted also for use in educational domain (mostly in MOOCs) to mediate communication between students and teachers. As students are only novices in topics they learn about, they may need various scaffoldings to achieve effective question answering. In this work, we focus specifically on automatic recommendation of tags classifying students' questions. We propose a novel method that can automatically analyze a text of a question and suggest appropriate tags to an asker. The method takes specifics of educational domain into consideration by a two-step recommendation process in which tags reflecting course structure are recommended at first and consequently supplemented with additional related tags.
Evaluation of the method on data from CS50 MOOC at Stack Exchange platform showed that the proposed method achieved higher performance in comparison with a baseline method (tag recommendation without taking educational specifics into account).
Recent research has provided solid evidence that emotions strongly affect motivation and engagement, and hence play an important role in learning. In BIG-AFF project, we build on the hypothesis that ``it is possible to provide learners with a personalised support that enriches their learning process and experience by using low intrusive (and low cost) devices to capture affective multimodal data that include cognitive, behavioural and physiological information''. In order to deal with the affect management complete cycle, thus covering affect detection, modelling and feedback, there is lack of standards and consolidated methodologies. Being our goal to develop realistic affect-aware learning environments, we are exploring different approaches on how these can be supported by either by traditional non-intrusive interaction sources or low intrusive and inexpensive sensing devices. In this work we describe the main issues involved in two user studies carried out with high school learners, highlight some open problems that arose when designing the corresponding experimental settings. In particular, the studies involved varied nature of information sources and each focused on one of the approaches. Our experience reflects the need to develop an extensive knowledge about the organization of this type of experiences that consider user-centric development and evaluation methodologies.
Effective exercise selection based on learner characteristics is important for Intelligent Tutoring Systems to improve learning. Based on a literature review, we categorize learner characteristics used for adaptation in an ITS. We then present a preliminary framework of the relationship between some of these learner characteristics, with an emphasis on personality, and how they can be used by an ITS to adapt exercise selection.
Personalized educational systems are able to provide learners questions of specified difficulty. Since learners differ, the appropriate level of difficulty may vary and it may be impossible to find an universal setting. We implemented a version of an adaptive educational system for geography practice that allows learners to adjust difficulty of questions. We evaluated this feature using a randomized control experiment. The overall results show only a small effect of the adjustment. A more detailed analysis, however, shows that for some groups of learners the effect can be important, although not necessarily advantageous. The collected data from the experiment provide insight into how to tune question difficulty automatically.
This paper presents current approaches and open issues regarding the modeling of users' physical activity when learning motor skills, such as those required to dance, play a musical instrument, practice sports or train in martial arts. On the one hand, it reveals the lack of personalized psychomotor learning systems and how the modeling of users' physical activity is just now becoming part of UMAP (User Modeling, Adaptation and Personalization) community research agenda. On the other hand, it proposes the Labanotation as a way for describing the movements performed during the users' physical activity, and comments on related works which show that it seems to be feasible to perform this labeling automatically with machine learning techniques. To touch down the proposal, the applicability of Labanotation for modeling the psychomotor activity when learning defensive martial arts movements such as those performed jointly in pairs in Aikido is analyzed.
Modern constructivist approaches to education dictate active experimentation with the study material and have been linked with improved learning outcomes in STEM fields. During classroom time we believe it is important for students to experiment with the lecture material since active recall helps them to start the memory encoding process as well as to catch misconceptions early and to prevent them from taking root. In this paper, we report on our experiences using ASQ, a Web-based interactive presentation tool in a functional and logic programming course taught at the Faculty of Informatics and Information Technologies at the Slovak University of Technology in Bratislava. ASQ allowed us to collect immediate feedback from students and retain their attention by asking complex types of questions and aggregating student answers in real time. From our experience we identified several requirements and guidelines for successfully adopting ASQ. One of the most critical concerns was how to estimate the time when to stop collecting the students' answers and proceed to their evaluation and discussion with the class. We also report the students' feedback on the ASQ system that we collected in the form of the standard SUS questionnaire.
Since 2007, the PATCH workshop series have been a gathering place for researchers and professionals from various countries and institutions to discuss the topics of digital access to Cultural Heritage and specifically the personalization aspects of this process. Due to this rich history, the reach of the PATCH workshop in various research communities is extensive. PATCH 2017 is another link in the long chain of PATCH events and we hope that it will point out future research challenges and directions and its success will pave the way to future events. Following the successful series of PATCH workshops, PATCH 2017 is organized as the meeting point between state of the art cultural heritage research and personalization -- using any kind of technology, while focusing on ubiquitous and adaptive scenarios, to enhance the personal experience in cultural heritage sites. The workshop is aimed at bringing together researchers and practitioners who are working on various aspects of cultural heritage and are interested in exploring the potential of state of the art of personalized approaches that may enhance the CH visit experience. In this edition we received 7 submissions, 2 full papers (28%), 3 short papers (42%), 1 demo paper (14%), and 1 position paper (14%). To select the workshop papers a peer-review process was carried out. At least three members of the Program Committee (which is listed below) were assigned to each paper. As result, we have accepted all the papers, and downgraded 1 short paper to position paper. The 2017 workshop includes contributions covering diverse research aspects, such as: advanced brain informatics and IoT approaches to understand museum visitors' behavior or to personalize their visits; novel models for information retrieval, information visualization and automated personalized content generation, social recommendation of CH information based on Linked Open Data; and a study of the interplay among human cognitive processing differences and cultural heritage activities towards gaming experience and performance. We believe that this is a nice spectrum of topics and we wish you to enjoy reading the workshop proceedings. The contributions collected in this workshop reflect these topics.
Digital museum guides - often together with eye trackers as innovative gadgets for intuitive interaction - provide attractive new ways for museums to communicate information to visitors and analyze their behaviour. In this paper, we investigate an approach to understand the gaze bedhaviour of persons viewing paintings in a museum. We present a method that can detect focussed areas (AOF) by analysing the fixation duration for the pixels of a painting. We can provide evidence that the viewing behaviour of laymen in a museum differs from what an expert expects according to the art historic relevance of certain regions of interest (ROI) in a painting. Consequently, museum educators have to apply intelligent assistance strategies that allow visitors to fully appreciate exhibits during their visit a of museum.
This demo presents a platform for the definition of IoT-enhanced visits to Cultural Heritage (CH) sites. The platform is characterized by an End-User Development paradigm applied to the Internet of Things technologies and customized for the CH domain. It allows different stakeholders to configure the behavior of smart objects in order to create more engaging visit experience and to increase the appropriation of CH content by visitors.
This article describes a recommender system (RS) in the cultural heritage area, which takes into account the activities on social media performed by the target user and her friends. For this purpose, the system exploits linked open data (LOD) as well. More specifically, the proposed RS (i) extracts information from social networks (e.g., Facebook) by analyzing content generated by users and those included in their social networks; (ii) performs disambiguation tasks through LOD tools; (iii) profiles the user as a social graph; (iv) provides the actual user with personalized suggestions of artistic and cultural resources by integrating collaborative filtering algorithms with semantic technologies for leveraging LOD sources such as DBpedia and Europeana.
There is a growing interests in integration of Internet of Things (IoT) in smart environments, which creates an opportunity to understand users' information needs using onsite physical sensor logs. However, the physical context creates numerous external factors that play a role in users' information interactions, thus creating new external biases in the collected information interaction logs. In order to provide an effective personalized experiences for users in smart environment, we need to take care of these external biases in the behavioral user models. Our general aim is to understand users' onsite physical behaviors for providing online and onsite personalized services like personalized tour guides. We focus on the cultural heritage domain and collect onsite users' physical information interaction logs of visits in a museum. This prompts the question: How to understand users' behavior in the existence of external biases? Our main finding is that users behave differently in their solitude in comparison to a busy museum situation. Specifically, visitors' crowd bias has a considerable effect on users' following position rank bias based check-in behavior. Our study investigates on understanding users' onsite physical behavior accurately, which can improve the state-of-the-art onsite behavioral user models.
Cultural heritage (CH) is an attractive domain for experimenting with novel technologies for various reasons. In general, the research focuses is on experimenting the potential of the novel technology, while having a high-quality content is necessary for experimentation in a realistic setting, but it is not the focus of the research. While generally ignored, it seems that automatic content creation is one of the main challenges for wide adoption of CH application in practice. Some effort was invested in automating the process without providing a real solution. It seems that recent semantic web techniques and large content digitization and standardization efforts pave the way for trying again to suggest ways for automatic creation of personalized and context aware coherent presentations from freely available content on the fly.
The exploration of cultural heritage information is challenged by the fact that most data provided by online resources is fragmented and it is repository or application-centered. In order to address this issue, a data integration approach should be adopted, that makes it possible to generate custom views, focused on the user's information needs, but easily extensible to support the inspection of topically related contents.
In this paper, we present a model supporting the management of thematic maps for information exploration, and their integration with query expansion during the interaction with the user. Our model is based on: (i) an ontological domain knowledge representation for describing the meaning of concepts and their semantic relations; (ii) a semantic interpretation model for identifying the concepts referenced in the user's queries. We are experimenting our model in the OnToMap Participatory GIS, which manages interactive community maps for information sharing and participatory decision-making.
Common design practices of current cultural heritage activities barely take into account the contextual, cultural, and cognitive characteristics of visitors. Bearing in mind that information processing is substantial in such activities, this paper investigates the interplay among human cognitive differences and cultural heritage gaming activities towards players' performance and visual behavior. Three user studies were conducted under the field dependence/independence theory, which underpin cognitive differences in visual perceptiveness and contextual information handling. Findings are expected to provide useful insights for practitioners and researchers with the aim to design playful cultural activities tailored to the users' cognitive preferences.
It is our great pleasure to welcome you to the UMAP 2017 Workshop on Surprise, Opposition, and Obstruction in Adaptive and Personalized Systems (SOAP). Following the successful first edition of the workshop at UMAP 2016, we are happy to see a continuing and increasing interest in the workshop's topics. As with the first edition, for the second edition we were able to accept four highly relevant submissions, allowing us to discuss the challenges of recommending unexpected, nonetheless relevant and impactful artifacts during a focused half-day workshop. With the workshop being originally motivated by interviews with music creators and producers who articulated a strong rejection of "more-of-the-same" search engines and recommender systems as they challenge their notion of originality and, ultimately, pose a threat to their artistic identity, we realized that a demand for adaptive and personalized systems that not only have the capability to surprise, but also to oppose and even obstruct can be found in a wider field. In fact, this coincides with ongoing trends to deal with and escape generally negatively connoted effects of automatic recommender systems, such as the so-called "filter-bubble". Apart from the potential dangers of such effects on the unreflecting user, there seems to be a growing impression that collaborative, as well as content-based recommender systems keep making obvious, uninspiring, and therefore disengaging suggestions based on previous interactions. Over the last years, this has emphasized the value of system qualities beyond pure accuracy, e.g., diversity, novelty, serendipity, or unexpectedness, to keep the user satisfied. In fact, these approaches to kicking the user out of his or her "comfort zone" seem to be highly promising methods to increase satisfaction with a system in the long run.
There is a long tradition in recommender systems research to evaluate systems using quantitative performance measures on fixed datasets. As a reaction to this narrow accuracy-based focus in research, novel qualities beyond pure accuracy are emphasized in recent research; among them are surprise and opposition.
This position paper considers that the perception of surprise and/or opposition may be purposely prepared when several recommendations are provided (e.g., in terms of a music playlist) or the user is given the choice between several options.
Altering users' perception and triggering according behavior is well rooted in research on priming from psychology and nudge theory from the field of economic behavior.
In this position paper, we propose how priming and nudging may be integrated into the design and evaluation of recommender systems to arouse surprise and opposition.
This paper presents a novel, graph embedding based recommendation technique. The method operates on the knowledge graph, an information representation technique alloying content-based and collaborative information. To generate recommendations, a two dimensional embedding is developed for the knowledge graph. As the embedding maps the users and the items to the same vector space, the recommendations are then calculated on a spatial basis. Regarding to the number of cold start cases, precision, recall, normalized Cumulative Discounted Gain and computational resource need, the evaluation shows that the introduced technique delivers a higher performance compared to collaborative filtering on top-n recommendation lists. Our further finding is that graph embedding based methods show a more stable performance in the case of an increasing amount of user preference information compared to the benchmark method.
In this paper, we propose a framework for computational serendipity. The framework is used in a recommender system context to find personalized serendipity and meanwhile stimulate user's curiosity. The framework is novel to the serendipity research community in that it decomposes the concept of serendipity into two elements: surprise and value; and provides computational approaches to modeling both of them. The framework also incorporates the concept of curiosity to keep users' interests over a long term. It brings together several fields including information retrieval, cognitive science, computational creativity in artificial intelligence, and text mining. We will describe the framework first and then evaluate it with an implementation called StumbleOn in the health news context. The evaluation serves as a proof-of-concept of this computational serendipity framework.
A music listener's mainstreaminess indicates the extent to which her listening preferences correspond to those of the population at large. However, formal definitions to quantify the level of mainstreaminess of a listener are rare and those available define mainstreaminess based on fractions between some kind of individual and global listening profiles. We argue, in contrast, that measures based on a modified version of the well-established Kullback-Leibler (KL) divergence as well as rank-order correlation coefficient may be better suited to capture the mainstreaminess of listeners. We therefore propose two measures adopting KL divergence and rank-order correlation and show, on a real-world dataset of over one billion user-generated listening events (LFM-1b), that music recommender systems can notably benefit when grouping users according to their level of mainstreaminess with respect to these two measures. This particularly holds for the frequently neglected listener group which is characterized by low mainstreaminess.
The importance of user modeling and personalization is taken for granted in several scenarios. According to this widespread paradigm, each user can be modeled through some (explicitly or implicitly gathered) information about her knowledge or about her preferences, in order to adapt the behavior of a generic intelligent system to her specific characteristics. However, the recent spread of social network and self-tracking devices has totally changed the rules for personalization. On one side, the spread of social network platforms radically changed and renewed many consolidated behavioral paradigms.
Thanks to the heterogeneous nature of the discussions that take place on social networks, a lot of new data are continuously available and can be gathered and exploited to build richer and more complete user models, to discover latent communities, to infer information about users' emotions and personality traits, and also to study very complex phenomena, such as those related to the psycho-social sphere, in a totally new way. At the same time, self-tracking devices are becoming more and more pervasive, and a plethora of personal data is today available by exploiting such tools.
These devices model and track a lot of signals that pure content-based information which is commonly spread on social networks can't actually handle. Reasoning on these data can enable predictions about the user's behavior, health, and goals. As a consequence, it is very important to think about a new generation of user models that are able to effectively merge the information coming from both information sources, while also taking into account the fact that user models evolve over time.
Daily physical activity not only empowers the body, but it also invigorates the mind and helps people cope with the struggle of everyday life. A balanced amount of moderate to vigorous physical activity is recommended. Major barriers that lead to low levels of physical activity are lack of time and motivation. The objective of this paper is to generate individual recommendations to improve physical activity by using if-then plans - so called Implementation Intentions. We developed a mobile application named DayActivizer to collect all the necessary activity data by the user. Based on the collected data, the application automatically recommends activities within if-then plans with an increasing degree of physical effort to counteract insufficient physical exercise concerning individual daily routine. To evaluate our approach, we conducted a field study (N=8) and qualitative interviews in which every participant was asked to examine the validity of the individual recommended implementation intentions.
With the increasing information overload, the identification of new users really relevant to the target user becomes more and more complicated. In this paper, we propose a social recommender based on a user model that takes into account not only her interests and preferences, but also their evolution over time and actual nature. To accurately assess the effectiveness of the proposed approach, over 1,600 users were monitored for a full year, thus collecting over 2,700,000 tweets. In this way, it was possible to deeply evaluate the proposed model, also through a comparative analysis with other state-of-the-art social recommender systems.
In human computer interaction, some of the user activities are intentional, and other unintentional, but user interfaces are usually designed to react only to intentional commands. However, user's unintentional activity contains many clues about a user, that can be beneficial to take into account in designing appropriate response. Current study focuses on these unintentional traces, that left behind by use of standard input devices, keyboard and mouse, and specifically, we try to predict users age and gender. Mouse and keyboard data used in this study, are collected in six different systems between 2011 and 2017 in total from 1519 subjects. Some supervised machine learning models yield to f-scores over 0.9 when predicted both user age or gender.
The use of social media, like Facebook, Twitter and LinkedIn, is nowadays very common and quite for sure each one of us has at least a digital profile on them. The information left of these platforms such as likes, posts, tweets and photos are very informative and can be used for deducting our preferences, tendencies and behaviors. The analysis of the social media footprints has become a relevant research topic in the last decade and many works have demonstrated how to extract some traits of the user's affective sphere. In this paper, we focus on the prediction of empathic tendencies of a subject as an index of the influence of emotions during decisional processes. This value can be included in the user profile and can be relevant in some scenarios, such as music and movie recommender systems, where the emotional component is strongly delineated. We propose an approach of empathy level prediction based on a linear regression algorithm over Facebook profiles. We use a word2vec representation of the textual contents of the user's time-line posts, a LDA and SVD vector representation of the user's likes and other general descriptive data. The evaluation performed has demonstrated the validity of the approach for predicting the empathy tendency and the results have showed some relevant correlations with some specific groups of user's descriptive features.
Virtual reality exergames have been demonstrated to provide high levels of exertion compared to traditional exercise, while players perceive less exertion. However, it is hard for people to be confident of whether they get the recommended levels of exercise. In this work, we present the "HappyFit" aesthetic interface that has been designed to be a pleasing ambient display that enables people to see their long-term user model of physical activity. The user model interface has been particularly designed to distinguish the source of the exercise and the user model is inferred from combinations of sensor data from worn devices that track steps and heart-rate. We show how "HappyFit" enables a person to gain an overview of the relative contributions of their exercise from both walking in daily life and playing virtual reality exergames. Our core contribution is the exploration of how to harness long term sensor data to build user models with aesthetic user interfaces that enable people to review and reflect on their physical activity.
It is our great pleasure to welcome you to the UMAP 2017 Fifty Shades of Personalization - Workshop on Personalization in Serious and Persuasive Games and Gameful Interactions. Serious games (games for purposes other than entertainment) as well as persuasive games (games for promoting desirable behavior without coercion) are increasingly adopted by scholars and have also found their way into industry. Elements of games are also increasingly used to design gameful interactions (this is also referred to as gamification). Serious and persuasive approaches focus on imparting knowledge and raising awareness about topics or issues, and also fostering attitude or behavior change in a desirable direction, for example towards a healthier lifestyle.
In an era when we are used to highly individualized, personal and ubiquitous interactions and with the possibility to collect an enormous amount of information about people's behaviors, habits and attitudes, personalization has increased much in significance since it became a topic in Human-Computer Interaction. Not only do we have advanced opportunities to personalize serious and persuasive games and gameful interactions, we have also scientific evidence that this is highly useful. Studies show that these technologies are more effective in educating users about certain topics and in supporting them in behavioral and attitudinal change, as well as in raising awareness and engaging them in specific topics, when they are personalized in contrast to employing a one-size-fits-all approach.
Although personalization of serious and persuasive games and gameful interactions is a vibrant and highly promising area and has become an important researched field, many aspects of it are still underexplored. Thus, there is common understanding on the importance of personalization itself, but also an ongoing debate and a growing number of research on the approaches used for personalization: Will we use subjective or objective variables for personalization? Will we use continuous (such as traits) or categorical (such as types) dimensions? Will we personalize according to specific interactions (e.g. game dynamics) or ends of the interaction (e.g. different goals)? Will we rely on an a priori personalization or will we be able to personalize in real-time? The various shades of personalization in serious games and gameful interactions will be the central aspect of the workshop and will form the basis for participants' discussions.
Obesity is a serious health problem that has been linked to the major cause of death worldwide, ischemic heart disease. There is a lot of research on influencing people to live healthier lives by being active and eating healthy foods. However, there is little research on influencing people to buy healthier foods at the point of sale especially online. Because people tend to cook and eat what they buy, making healthier choices when grocery shopping online could lead to healthier eating habits for consumers. To advance research in this area, we propose a framework that uses gamification elements to influence consumers to purchase healthier foods in e-commerce. In this position paper, we present our proposed framework and describe the implementation of some of the influence strategies and game design elements such as rewards, personalization, suggestion, self-monitoring and feedback. This paper contributes to the area of game design by describing possible guidelines that could lead to healthier food shopping habits for e-commerce consumers.
Our wider research project investigates the design of a persuasive game for preventing mental health problems and improving subjective wellbeing in a student population. In this paper, we explore how persuasive game elements and interactions can be adapted to different student personalities, active stressors and attitudes. In six focus groups we investigated (1) which key stressors are experienced by students, (2) what characteristics of students need to be considered for adapting game interactions and challenges, and (3) which approaches to personalisation could be applied. Participants were shown stories about a fictional student, conveying high and low levels of three personality traits (Conscientiousness, Emotional Stability and Extraversion), levels of active stressors, and varying attitudes towards change. Participants discussed how to tailor game interactions, activities and challenges to the characteristics of the fictional student. In general, participants perceived real-time personalisation using implicit measures as more effective, but recognised explicit profiling as a valuable complementary method. These findings have implications for the personalisation and design of persuasive game based interventions for health.
Gamification in the era of chatbots is a novel way to engage users with the chatbot application. When developing a gamified chatbot system, there are factors related to user types (ages, gender and others) that we should consider to effectively integrate the game elements into the chatbot while targeting the right audience. In this study, we discuss the development of an educational chatbot game 'CiboPoli', that's specialised in teaching children about healthy lifestyle through an interactive social game environment. The presented game is based on a paper prototype that we developed to teach primary school students about healthy diet and food waste management. The current approach will be more engaging and pose AI capabilities. This is still a work in progress and we plan to improve its design by incorporating additional components, such as dialog management module, user-specific knowledge module or machine learning module. Future work will be devoted to integrating machine learning to automatically identify learners emotions and provide personalised suggestions. Moreover, we tested the initial prototype with school students and found that it outperforms the paper version. Future work will focus on applying it to other domains and demographics.
Personalizing interactive systems including games increases their effectiveness. This paper explores and compares two main approaches to personalization: system-controlled and user-controlled adaptation. The results of large-scale exploratory studies of 1768 users show that both techniques to personalizing systems share seven common strengths of increasing users' perception of a system's relevance, usefulness, interactivity, ease of use, credibility and trust, and also increases users' self-efficacy. The results also reveal some unique strengths and weaknesses peculiar to each of the approaches that designers should take into consideration when deciding on a suitable adaptation technique to use in personalizing their systems. Users prefer system- over user-controlled adaptation.
Research has shown that Competition is a powerful intrinsic motivator of behavior change. However, little is known about the predictors of its persuasiveness and the moderating effect of culture. In this paper, we investigate the predictors of "the persuasiveness of Competition" (i.e. Competition) using three social influence con-structs: Reward, Social Comparison and Social Learning. Using a sample of 287 participants, comprising 213 individualists and 74 collectivists, we explored the interrelationships among the four social influence constructs and how the two cultures differ and/or are similar. Our global model, which accounts for 46% of the variation in Competition, reveals that Reward has the strongest influence on Competition, followed by Social Comparison. However, the model shows that Social Learning has no significant influence on Competition. Finally, our multigroup analysis reveals that, for our population sample, culture does not moderate the interrelationships among the four constructs. Our findings suggest that designers of gamified applications can employ Reward, Social Comparison and Competition as co-persuasive strategies to motivate behavior change for both cultures, as the susceptibilities of users to Reward and Social Comparison are predictors of their susceptibility to Competition.
Gamification has been used in a variety of application domains to promote behaviour change. Nevertheless, the mechanisms behind it are still not fully understood. Recent empirical results have shown that personalized approaches can potentially achieve better results than generic approaches. However, we still lack a general framework for building personalized gameful applications. To address this gap, we present a novel general framework for personalized gameful applications using recommender systems (i.e., software tools and technologies to recommend suggestions to users that they might enjoy). This framework contributes to understanding and building effective persuasive and gameful applications by describing the different building blocks of a recommender system (users, items, and transactions) in a personalized gamification context.
Exercise is essential for health and well-being. However, it can be difficult for people to meet the recommended amount of daily exercise simply due to the lack of motivation. It has recently become apparent that virtual reality games, even though they were not explicitly designed for exercise, have the potential to provide enough exercise to achieve recommended levels of activity for a day, while keeping people motivated. However, as these games have not generally not been designed for exercise, there is a risk that people may either under- or over-exert themselves. Therefore, in this paper we present and discuss our design for a virtual reality exergame that utilizes a user model and dynamic difficulty adjustment to deliver personalized activity levels and experiences.