Detailed Schedule

Abstract: Recommender Systems (RSs) help users search large amounts of contents by allowing them to identify the items that are likely to be more attractive or useful. RSs play an important role in many domains (e.g., e-commerce, e-tourism, social networks, entertainment), as they can potentially augment the users’ trust towards an application and orient their decisions or actions towards specific directions. The goal of this tutorial is to give participants a solid background of how to design and evaluate RSs, with a particular focus on user experience aspects, and to provide pragmatic guidelines to perform these activities more effectively. The tutorial is structured into three parts, and the teaching style will be “by-examples”. In the first part, after a general overview of recommender systems, we will analyze their design issues and will analyze different definitions of design quality. In the second part we will analyze “off-line” (system-centric) evaluation techniques. We will describe different quality metrics and how to measure them. We will highlight the benefits and pitfalls of off-line evaluation. The third part will explore “on-line” (user–centric) quality evaluation methodologies, focusing on the users’ perceived values and how these affect their behavioral intentions. We will describe in some details a unifying user-centric evaluation framework, called ResQue (Recommender systems’ Quality of user experience), and psychometric methods, including structural equation modeling (SEM).

Abstract: The role of recommender systems as a fundamental utility for electronic commerce and information access is well established with many commercially-available recommender systems providing benefits to both users and businesses. But, recommender systems tend to use simplistic user models that are additive in nature: new user preferences are simply added to the existing profiles. This additive approach ignores the notion of “situated action,” that is, the fact that users interact with systems within a particular context and items relevant within one context may be irrelevant in another. Little agreement exists among researchers as to what constitutes context, but its importance seems undisputed. In psychology, a change in context during learning has been shown to have an impact on recall. Research in linguistics has shown that context plays the important role of a disambiguation function. More recently, the role of context has been explored in intelligent information systems. In particular, a variety of approaches and architectures have emerged for incorporating context or situational awareness in the recommendation process.
The goal of this tutorial is to provide a broad overview of the problem of contextual recommendation and some of the recent solutions to the problem of modeling context. I will specifically focus on several approaches for integrating context in user modeling for personalized recommendation, including an approach inspired by a model of human memory and emphasizes the modeling of context based on observations of user behavior; another that emphasizes the role of domain knowledge and semantics as an integral part of user context, and finally, an approach that exploits social annotations, such as collaborative tagging, as the basis for inferring context.

Abstract: One of the major innovations in personalization in the last 20 years was the injection of social knowledge into the model of the user. The user is not considered an isolated individual any more, but a member of one or more communities. User communities have been facilitated by the striking advancements of electronic communications and in particular the penetration of the Web into people’s everyday routine. Communities arise in a number of different ways. Social networking tools typically allow users to proactively connect to each other. Alternatively, data mining tools discover communities of connected Web sites or communities of Web users. In this tutorial, the focus is on the latter type of community, which is commonly mined from logs of users’ activity on the Web. In the tutorial we will recall how this process has been used to model the users’ interests and personalize Web applications. On this basis, we will examine the effect of recent developments on the Web and particularly the advent of the social Web. We will explain how this development draws together the different viewpoints on Web communities and introduces new opportunities for community-based personalization. In particular, we will analyse the concept of active user community and show how this relates to recent efforts on mining social networks and media.

Title: Link, Like, Follow, Friend:The Social Element in User Modeling and Adaptation

Abstract: The social web is having a clear impact in our field of user modeling and adaptation. ‘Links’ and ‘Likes’ as well as ‘Followers’ and ‘Friends’ are part of a large source of data that is generated by users themselves, often for different purposes, and that provides an unprecedented potential for systems to understand their users and to adapt based on that understanding. As we can see from researchers and projects in a number of relevant fields, data on various manifestations of what users do socially on the web brings new opportunities. Exciting ideas are generated and first explorations show promising results. In this talk we take a look back at recent proposals and studies that consider the social web. We determine interesting patterns and we aim to understand the impact on methods and techniques for user modeling and adaptation. At the same time, the social web brings even more challenges. We look forward by identifying challenges that can drive our research. From technical challenges to explore the different social web sources to social challenges to understand how users behave when this potential is unlocked.

Abstract: Rating prediction is a well-known recommendation task aiming to predict a user’s rating for those items which were not rated yet by her. Predictions are computed from users’ explicit feedback, i.e. their ratings provided on some items in the past.
Another type of feedback are user reviews provided on items which implicitly express users’ opinions on items. Recent studies indicate that opinions inferred from users’ reviews on items are strong predictors of user’s implicit feedback or even ratings and thus, should be utilized in computation. As far as we know, all the recent works on recommendation techniques utilizing opinions inferred from users’ reviews are either focused on the item recommendation task or use only the opinion information, completely leaving users’ ratings out of consideration. The approach proposed in this paper is filling this gap, providing a simple, personalized and scalable rating prediction framework utilizing both ratings provided by users and opinions inferred from their reviews. Experimental results provided on a dataset containing user ratings and reviews from the real-world Amazon Product Review Data show the effectiveness of the proposed framework.

Abstract: Content recommender systems have become an invaluable tools in online communities where a huge volume of content items are generated for users to consume, making it difficult for users to find interesting content. Many recommender systems leverage articulated social networks or profile information (e.g, user background, interest, etc.) for content recommendation. These recommenders largely ignore the implied networks defined through user inter-actions. Yet these play an important role in formulating users’ common interests. We propose an interaction based content recommender which leverages implicit user interactions to determine the relationship trust or strength, generating a richer, more informed implied network. An offline analysis on a 5000 per-son, 12 week dataset from an online community shows that our approach out-performs algorithms which focus on articulated networks that do not consider relationship trust or strength.

Abstract: In academic studies, the evaluation of recommender system (RS) algorithms is often limited to offline experimental designs based on historical data sets and metrics from the fields of Machine Learning or Information Retrieval. In real-world settings, however, other business-oriented metrics such as click-through-rates, customer retention or effects on the sales spectrum might be the true evaluation criteria for RS effectiveness. In this paper, we compare different RS algorithms with respect to their tendency of focusing on certain parts of the product spectrum. Our first analysis on different data sets shows that some algorithms — while able to generate highly accurate predictions — concentrate their top 10 recommendations on a very small fraction of the product catalog or have a strong bias to recommending only relatively popular items than others. We see our work as a further step toward multiple-metric offline evaluation and to help service providers make better-informed decisions when looking for a recommendation strategy that is in line with the overall goals of the recommendation service.

Abstract: Often one needs to form teams in order to perform a complex collaborative task. Therefore, it is interesting and useful to assess how well constituents of a team have performed, and leverage this knowledge to guide future team formation. In this work we propose a model for assessing the reputation of participants in collaborative teams. The model takes into account several features such as the different skills that a participant has and the feedback of team participants on her/his previous works. We validate our model based on synthetic datasets extrapolated from real-life scenarios.

Abstract: Social web provides rich content for gaining an understanding about the users which can empower adaptation. There is a current trend to extract user profiles from social media content using semantic augmentation and linking to domain ontologies. The paper shows a further step in this research strand, ex-ploiting semantics to get a deeper understanding about the users by extracting the domain regions where the users focus, which are defined as viewpoints. The paper outlines a formal framework for extracting viewpoints from semantic tags associated with user comments. This enables zooming into the viewpoints at different aggregation layers, as well as comparing users on the basis of the areas where they focus. The framework is applied on YouTube content, illustrating an insight into emotions users refer to in their comments on job interview videos.

Abstract: The social graphs of Twitter users often overlap, such that retweets may cause duplicate posts is a user’s incoming stream of tweets. Hence, it is important for the retweets to strike the balance between sharing information and flooding the recipients with redundant tweets. In this work, we present an exploratory analysis that assesses the degree of duplication caused by a set of real retweets. The results of the analysis show that although the overall duplication is not severe, high degree of duplication is caused by tweets of users with a small number of followers, which are retweeted by users with a small number of followers. We discuss the limitations of this work and propose several enhancements that we intend to pursue in the future.

Abstract: Tagging has become a powerful means for users to find, organize, understand and express their ideas about online entities. However, tags present great challenges when researchers try to incorporate them into the prediction task of recommender systems. In this paper, we propose a novel approach to infer user preference from tag relevance, an indication of how strong each tag applies to each item in recommender systems. We also present a methodology to choose tags that tell most about each user’s preference. Our preliminary results show that at certain levels, some of our algorithms perform better than previous work.

Abstract: Recommending products to new buyers is an important problem for online shopping services, since there are always new buyers joining a deployed system. In some recommender systems, a new buyer will be asked to indicate her/his preferences on some attributes of the product (like camera) in order to address the so called cold-start problem. Such collected preferences are usually not complete due to the user’s cognitive limitation and/or unfamiliarity with the product domain, which are called partial preferences. The fundamental challenge of recommendation is thus that it may be difficult to accurately and reliably find some like-minded users via collaborative filtering techniques or match inherently preferred products with content-based methods. In this paper, we propose to leverage some auxiliary data of online reviewers’ aspect-level opinions, so as to predict the buyer’s missing preferences. The resulted user preferences are likely to be more accurate and complete. Experiment on a real user-study data and a crawled Amazon review data shows that our solution achieves better recommendation performance than several baseline methods.

Abstract: Most of the research studies on recommender systems are focused on single-domain recommendations. With the growth of multi-domain internet stores such as iTunes, Google Play, and Amazon.com, an opportunity to offer recommendations across different domains become more and more attractive. But there are few research studies on cross-domain recommender systems. In this paper, we study both the cold-start problem and the hypothesis that cross-domain recommendations provide more accuracy using a large volume of user data from a true multi-domain recommender service. Our results indicate that cross-domain collaborative filtering could significantly improve the quality of recommendation in cold start context and the auxiliary profile size plays an important role in it.

Abstract: Several recommender systems have been proposed in the literature for adaptively suggesting useful references to researchers with different interests. However, in order to access the knowledge contained in the recommended papers, the users need to read the publications for identifying the potentially interesting concepts. In this work we propose to overcome this limitation by utilizing a more semantic approach where concepts are extracted from the papers for generating and explaining the recommendations. By showing the concepts used to find the recommended articles, users can have a preliminary idea about the filtered publications, can understand the reasons why the papers were suggested and they can also provide new feedback about the relevance of the concepts utilized for generating the recommendations.

Abstract: We show that people’s travel destinations are predictable based on simple features of their home and destination. Using geotagged Twitter data from over 200,000 people in the U.S., with a median of 10 visits per user, we use machine learning to classify whether or not a person will visit a given location. We find that travel distance is the most important predictive feature. Ignoring distance, using only demographic features pertaining to race, age, income, land area, and household density, we can predict travel destinations with 84% accuracy. We present a careful analysis of the power of individual and grouped demographic features to show which ones have the most predictive impact for where people go.

Abstract: We develop and test algorithms for assessing the similarity of a person’s days based on location traces recorded from GPS. An accurate similarity measure could be used to find anomalous behavior, to cluster similar days, and to predict future travel. We gathered an average of 46 days of GPS traces from 30 volunteer subjects. Each subject was shown random pairs of days and asked to assess their similarity. We tested eight different similarity algorithms in an effort to accurately reproduce our subjects’ assessments, and our statistical tests found two algorithms that performed better than the rest. We also successfully applied one of our similarity algorithms to clustering days using location traces.

Abstract: This paper presents our work on supporting scrutable user models for use in mobile applications that provide personalised item recommendations. In particular, we describe a mobile lifestyle application in the fine-dining domain, designed to recommend meals at a particular restaurant based on a person’s user model. The contributions of this work are three-fold. First is the mobile application and its personalisation engine for item recommendation using a content and critique-based hybrid recommender. Second, we illustrate the control and scrutability that a user has in configuring their user model and browsing a content list. Thirdly, this is validated in a user experiment that illustrates how new digital features may revolutionise the way that paper-based systems (like restaurant menus) currently work. Although this work is based on restaurant menu recommendations, its approach to scrutability and mobile client-side personalisation carry across to a broad class of commercial applications.

Abstract: Personalized systems are gaining popularity in various mobile scenarios. In this work, we take on the challenges associated with the automotive domain and present a user-adaptive graphical renderer. By supporting a strictly model-based development processes, we meet the rigid requirements of the industry. The proposed architecture is based on the UIML standard and a novel rule-based adaptation framework.

Abstract: We present an emotion model based on social tags, which is built upon an automatically generated lexicon that describes emotions by means of synonym and antonym terms. Using this model we develop a number of methods that transform social tag-based item profiles into emotion-oriented item profiles. We show that the model’s representation of a number of basic emotions is in accordance with the well known psychological circumplex model of affect, and we report results from a user study that show a high precision of our methods to infer the emotions evoked by items in the movie and music domains.

Abstract: This paper demonstrates that a user of multilingual search has different interests depending on the language used, and that the user model should reflect this. To demonstrate this phenomenon, the paper proposes and evaluates a set of result re-ranking algorithms based on various user model representations.

Abstract: In our connected workplaces it can be hard to work calm and focused. In a simulated work environment we manipulated the stressors time pressure and email interruptions. We found effects on subjective experience and working behavior. Initial results indicate that the sensor data that we collected is suitable for user state modeling in stress related terms.

Abstract:Affective support can be provided through personalized recommendations integrated within learning management systems (LMS). We have ap-plied the TORMES user centered engineering approach to involve educators in a recommendation elicitation process in a distance learning (DL) context.

Abstract: Finding satisfactory scientific literature is still a very time-consuming task. In the last decade several tools have been proposed to approach this task, however only few of them actually analyse the whole document in order to select and present it to the user and even less tools offer any kind of explanation of why a given item was retrieved/recommended.
The main goal of this demonstration is to present the RES system, a tool intended to overcome the limitations of traditional recommender and personalized information retrieval systems by exploiting a more semantic approach where concepts are extracted from the papers in order to generate and then explain the recommendation. RES acts like a personalized interface for the well-known CiteSeerX system, filtering and presenting query results accordingly to individual user’s interests.

Abstract: The main contribution of this work is the comparison of different techniques for representing user preferences extracted by analyzing data gathered from social networks, with the aim of constructing more transparent (human-readable) and serendipitous user profiles. We compared two different user models representations: one based on keywords and one exploiting encyclopedic knowledge extracted from Wikipedia. A preliminary evaluation involving 51 Facebook and Twitter users has shown that the use of an encyclopedic-based representation better reflects user preferences, and helps to introduce new interesting topics.

Abstract: We propose a novel approach to personal ontologies, grounded on the concept of affordance and on the ontological theory of Von Uexküll, in which each concept can be viewed under different perspectives depending on the subjectivity of the user and thus can yield tailored semantic relationships or properties. We suggest a cognitive middle-layer interface between the user and the ontology, which is able on the run to modify and adapt the ontology to the user needs. The goal is to obtain an adapted version of the ontology that is tailored both to the context and to the user prospective and expertise, without the need of explicitly maintaining a high number of ontologies.

Abstract: This demo presents an approach to recognising and interpreting social cues-based interactions in computer-enhanced job interview simulations. We show what social cues and complex mental states of the user are relevant in this interaction context, how they can be interpreted using static Bayesian Networks, and how they can be recognised automatically using state-of-the-art sensor technology in real-time.

Abstract: This paper briefly introduces Topolor, a social personalized adaptive e-learning system, which aims at improving fine-grained social interaction in the learning process in addition to applying classical adaptation based on user modeling. Here, we present the main features of Topolor and its preliminary evaluation that showed high system usability from a student’s perspective. The intention is to demonstrate Topolor hands-on at the conference.

Abstract: When purchasing an interactive product, users nowadays seek more than a flawless functionality and a comfortable ease of use. Products need to be enjoyable and exciting to have a unique selling point. User Experience (UX) is constituted by the instrumental qualities as well as the hedonic qualities of a product and impacts on the user’s overall appraisal. One way to improve product appraisal is the use of surprise as a design element. Surprising product design has been shown to be beneficial for the user and the rating of a product. By using the classical computer game Tetris, the impact of surprise on UX ratings of a digital, interactive computer game was investigated. The results of our study stress two points. First, unexpected events with undesirable consequences lead to negative surprises which in turn impede users’ information processing and have a bad impact on user experience. Second, whether unexpected events with desirable consequences lead to positive surprises, mainly depends on the interaction context and on the kind of system under consideration.

Abstract: In this paper we describe preliminary approaches for content-based recommendation of Pinterest boards to users. We describe our representation and features for Pinterest boards and users, together with a supervised recommendation model. We observe that features based on latent topics lead to better performance than features based on user- assigned Pinterest categories. We also nd that using social signals (repins, likes, etc.) can improve recommendation quality.

Abstract: In this paper we explore the use of semantics to improve diversity in recommendations. We use semantic patterns extracted from Linked Data sources to surface new connections between items to provide diverse recommendations to the end users.We evaluate this methodology by adopting a bottom-up approach, i.e. we ask users of a crowdsourcing platform to choose a movie recommendation from among ve options. We evaluate the results in terms of a diversity measure based on the semantic distance of topics and genres of the result list. The results of the experiment indicate that there are features of semantic patterns that can be used as an indicator of its suitability for the recommendation process.

Abstract: Research on user modeling based on social network information has shown that some user characteristics can be accurately inferred from users’ digital traces. This kind of information can be used to inform user models of adaptive systems for personalizing the system. This paper addresses a crucial question for practical application of this approach: Are users actually willing to provide their social Web profiles and how do they perceive this? An empirical study conducted with medical students shows that although participants are using social networks, they are reluctant about providing their identities and consider these portals rather private. The outcomes of the study uncover a clear need for further research on enhanced privacy and enhanced trust.

Abstract: This paper describes a rst investigation of potential domain expertise in Pinterest. We introduce measures for characterizing the volume and coherence of Pinterest users’ pinning activity in a given category, their perceived and declared category-specic expertise and the response from the social network. We use such signals in the context of a supervised ML framework and report encouraging preliminary results on the task of mining potential experts for 4 popular content categories.

Abstract: In this paper, we describe a methodology for capturing player experience while interacting with a game and we present a data-driven approach for modeling this interaction. We believe the best way to adapt games to a specific player is to use quantitative models of player experience derived from the in-game interaction. Therefore, we rely on crowd-sourced data collected about game context, players behavior and players self-reports of dierent aective states. Based on this information, we construct estimators of player experience using neuroevolutionary preference learning. We present the experimental setup and the results obtained from a recent case study where accurate estimators were constructed based on information collected from players playing first-person shooter game. The framework presented is part of a bigger picture where the generated models are utilized to tailor content generation to particular player’s needs and playing characteristics.

Abstract: Libraries have large growing book collections and users have difficulty in browsing the whole collection when choosing new books to read, particularly when looking for books without a dened goal. In this case, recommendation systems are useful and play an important role in improving library usability. Recommendations are based on ratings and the quality of recommendations depends on the quality of the ratings. Studies show that users rate more items if scales have smaller granularity. In this paper, we propose a dierent rating scale for the book recommendation scenario in a collaborative ltering set-up and study how time influences rating relevance. Our ndings suggest that the collaborative ltering algorithm benets from a rating scale with smaller granularity. Moreover, if some conditions are met, rating prediction quality can be improved if we give lower weight to older ratings

Abstract: We compared three term scoring methods in their ability to extract descriptive terms from a knowledge worker’s document collection. We compared the methods in two dierent evaluation scenarios, both from the perspective of the user: a per-term evaluation, and a holistic (term cloud) evaluation. We found that users tend to prefer a term scoring method that gives a higher score to multi-word terms than to single-word terms. In addition, users are not always consistent in their judgements of term proles, if they are presented in dierent forms (as list or as cloud).

Abstract: It is difficult for publishers to include the right links in documents, because they cannot predict all actions their users might want to perform. Existing adaptive navigation systems can generate relevant links, but doing this on a Web scale is non-trivial, especially if the targets are dynamic actions. As a result, adaptation often happens in a centralized way on a limited or closed document and action set. Distributed affordance is a technology to automatically generate links from any Web resource to matching actions from an open set of Web services, based on semantic annotations. In this paper, we indicate how this technology can be applied to adaptive navigation. We investigate how the generated links can be represented and how their relevance can be guaranteed. Based on that, we conclude that semantic technologies are an enabler to perform adaptive navigation to dynamic actions in a distributed way.

Abstract: This project poster introduces the recently started EEXCESS project, which aims at Enhancing Europes eXchange in Cultural Educational and Scientific Resource. Europe has digitised vast amounts of cultural, scientific and educational content like for example scientific research, historical sound recordings, images of sculptures, lms and sheet music. However, since content dissemination on the Web is driven by a small number of large central hubs like social networks or search engines, this cultural and scientic treasures has hardly been recognized by the general public or utilized in scientic and educational processes. EEXCESS aims to develop personalized and contextualized recommendation technologies to augment existing content dissemination channels (e.g. social media) and content creation process (e.g. blogging) for distributing high-quality educational, scientic and cultural content. In this project poster we present the underlying idea and related work with focus on user modeling and personalized, context-aware recommendation.

Abstract: Linked Data principles allow for easy discovery, reference, access and reuse ofWeb data. The user modeling community already widely exploits SemanticWeb technologies, but the Linked Data approach is still not widely adopted. The LinkedUp project aims to advance the exploitation of open data on the Web, particularly for education. In this paper, we discuss the relevance of Linked Data for user modeling and personalization, and how to participate in and prot from the various initiatives of LinkedUp.

Abstract: Traditional recommender system evaluation focuses on raising the accuracy, or lowering the rating prediction error of the recommendation algorithm. Recently, however, discrepancies between commonly used metrics (e.g. precision, recall, root-mean-square error) and the experienced quality from the users’ have been brought to light. This project aims to address these discrepancies by attempting to develop novel means of recommender systems evaluation which encompasses qualities identified through traditional evaluation metrics and user-centric factors, e.g. diversity, serendipity, novelty, etc., as well as bringing further insights in the topic by analyzing and translating the problem of evaluation from an Information Retrieval perspective.

Abstract: techplay.mobi is a research project aimed at building a platform for developing mobile serious games. The novelty is to support the design of games that consider affective features during the game interaction, provide personal-ized responses according to users’ interactions, comply with accessibility re-quirements and focus on improving psycho-educational competences and on promoting critical thinking.

Title: Language Adaptation

Abstract: As we all know, more and more of life is now manifested online, and many of the digital traces that are left by human activity are increasingly recorded in natural-language format. This availability offers us the opportunity to glean user-modeling information from individual users’ linguistic behaviors. This talk will discuss the particular phenomenon of individual language adaptation, both in the short term and in the longer term. We’ll look at connections between how people adapt their language to particular conversational partners or groups, on the one hand, and on the other hand, those people’s relative power relationships, quality of relationship with the conversational partner, and propensity to remain a part of the group.

Abstract: This paper studies the effect of individual differences in human cognition on user performance in authentication tasks. In particular, a text-based password and a recognition-based graphical authentication mechanism were deployed in the frame of an ecological valid experimental design, to investigate the effect of individuals’ different cognitive processing abilities toward efficiency and effectiveness of user authentication tasks. A total of 107 users participated in the reported study during a three-month period between September and November 2012. The results of this recent study can be interpreted under the light of human information processing as they demonstrate a main effect of users’ cognitive processing abilities on both efficiency and effectiveness related to authentication mechanisms. The main findings can be considered valuable for future deployment of adaptive security mechanisms since it has been initially shown that specific cognitive characteristics of users could be a determinant factor for the adaptation of security mechanisms.

Information graphics (bar charts, line graphs, etc.) in popular media generally have a high-level message that they are intended to convey. These messages are seldom repeated in the document’s text yet contribute to understanding the overall document. The relative perceptual effort required to recognize a particular message is a communicative signal that serves as a clue about whether that message is the one intended by the graph designer. This paper presents a model of relative effort by a viewer for recognizing different messages from grouped bar charts. The model is implemented within the ACT-R cognitive framework and has been validated by human subjects experiments. We also present a statistical analysis of the contribution of effort in recognizing the intended message of a grouped bar chart.

Abstract: This paper investigates the possibility of using information from brain signals, obtained through a light and inexpensive Brain Computer Inter-face (BCI), in order to dynamically adjust the difficulty of an educational video game and adapt the level of challenge to players’ abilities. In this experiment, attention levels of Tetris players – measured with the BCI – have been evaluated as a function of game difficulty. Processing of the data revealed that both in intra- and inter- player analysis, an increase in game difficulty was followed by an increase in attention. These results come in accordance with similar experiments performed with a 19 sensor EEG cap, as opposed to the single-dry-sensor BCI used here. These findings give new possibilities in the development of educational games that adapt to the mental state of player/learner.

Abstract: Due to the ageing of the world population, a growing number of elderly people remain in their homes, requiring different levels of care. Our formative user studies show that the main concern of elderly people and their families is “fall detection and safe movement in the house”, while eschewing intrusive monitoring devices. This paper introduces a statistical model based on non-intrusive sensor observations that posits whether a person is {\em not} safe by identifying unusually long periods of inactivity within different regions in the home. Evaluation on two real-life datasets shows that our system outperforms a state-of-the-art system.

People with dyslexia often face huge writing difficulties. Spell-checkers/predictors can help, but the current systems are not appropriate for them, because of the assumptions behind the models and because of heavy-to-use interfaces. This paper presents a system for spellchecking/predicting words, which can adapt both its model and its interface according to the individual behavior. The model takes into account typical errors made by people with dyslexia, such as boundary errors, and the context for correcting real-word errors. The interface aims at reducing interaction with the user. The model and the interface are easily adaptable to general use.

Abstract: Data about users is rapidly growing, collected by various online applications and databases. The ability to share user data across applications can offer benefits to user in terms of personalized services, but at the same time poses privacy risks of disclosure of personal information. Hence, there is a need to ensure protection of user privacy while enabling user data sharing for desired personalized services. We propose a policy framework for user data sharing based on the purpose of adaptation. The framework is based on the idea of a market, where applications can offer and negotiate user data sharing with other applications according to an explicit user-editable and negotiable privacy policy that defines the purpose, type of data, retention period and price.

Abstract: Context-aware recommender systems (CARS) adapt their recommendations to users’ specific situations. In many recommender systems, particularly those based on collaborative filtering, the contextual constraints may lead to sparsity:
fewer matches between the current user context and previous situations. Our earlier work proposed an approach called differential context relaxation (DCR), in which different subsets of contextual features were applied in different components of a recommendation algorithm. In this paper, we expand on our previous work on DCR, proposing a more general approach — differential context weighting (DCW), in which contextual features are weighted. We compare DCR and DCW on two real-world datasets, and DCW demonstrates improved accuracy over DCR with comparable coverage. We also show that particle swarm optimization (PSO) can be used to efficiently determine the weights for DCW.

Abstract: Context-aware recommender systems aim at outperforming tradition-al context-free recommenders by exploiting information about the context under which the users’ ratings are acquired. In this paper we present a novel contextual pre-filtering approach that takes advantage of the semantic similarities be-tween contextual situations. For assessing context similarity we rely only on the available users’ ratings and we deem as similar two contextual situations that are influencing in a similar way the user’s rating behavior. We present an ex-tensive comparative evaluation of the proposed approach using several contextually-tagged ratings data sets. We show that it outperforms state-of-the-art context-aware recommendation techniques.

People-to-people recommendation differ from item recommendations in a number of ways, one of which is that individuals add information to their profile which is often critical in determining a good match. The most critical information can be in the form of free text or personal tags. We explore text-mining techniques to improve classical collaborative filtering methods for a site aimed at matching people who are looking for expert advice on a specific topic. We compare results from a LSA-based text similarity analysis, a simple user-user collaborative filter, and a combination of both methods used to recommend people to meet for a knowledge-sharing website. Evaluations show that LSA similarity has a better precision at low recall rates, whereas collaborative filters have a better precision at higher recall rates. A combination of both can outperform the results of the simpler algorithms.

Abstract: This paper uses a User-as-Wizard approach to evaluate how people apply diversity to a set of recommendations. In particular, it considers how diversity is applied for a recipient with high or low Openness to Experience, a personality trait from the Five Factor Model. While there was no effect of the personality trait on the degree of diversity applied, there seems to be a trend in the way in which it was applied. Maximal categorical diversity (across genres) was more likely to be applied to those with high Openness to Experience, at the expense of maximal thematic diversity (within genres).

Abstract: Email is a ubiquitous communication tool and constitutes a significant portion of social interactions. In this paper, we attempt to infer the personality of users based on the content of their emails. Such inference can enable valuable applications such as better personalization, recommendation, and targeted advertising. Considering the private and sensitive nature of email content, we propose a privacy-preserving approach for collecting email and personality data. We then frame personality prediction based on the well-known Big Five personality model and train predictors based on extracted email features. We report prediction performance of 3 generative models with different assumptions. Our results show that personality prediction is feasible, and our email feature set can predict personality with reasonable accuracies.

Title: Audience Selection in Computational Advertising

Abstract: Online interaction is becoming increasingly individualized both via explicit means such as customizations, options, add-ons, skins, apps, etc., and via implicit means, that is, large scale analyses of user behavior (individually and in aggregate) that allow automated, user-specific content and experiences such as individualized top news selection based on inferred interests, personal “radio stations” that capture idiosyncratic tastes from past choices, individually recommended purchases via collaborative filtering, and so on.
Not surprisingly, since online content and services are often ad-funded, online advertising is becoming increasingly user-specific as well, supported by the emerging discipline of Computational Advertising whose main goal is to find the “best match” between a given user in a given context and a suitable ad. There is a wide variety of users and contexts and the number of potential ads might be in the billions. Thus, depending on the definition of “best match” this problem leads to a variety of massive optimization and search problems, with complicated constraints.
The focus of this talk is audience selection, a form of advertising whereby advertisers specify the features of their desired audience, either explicitly, by specifying characteristics such as demographics, location, and context, or implicitly by providing examples of their ideal audience. A particular form of audience selection is interest-based advertising where the desired audience is characterized by its manifested interests. We will discuss how audience selection fits the optimization framework above, present some of the technical mechanisms for selection, and briefly survey some issues surrounding advertising and user privacy. We will conclude with some speculations about the future of online advertising and pertinent areas of scientific research.

Abstract: In recent years, models of student inquiry skill have been developed for relatively tightly-scaffolded science simulations. However, there is an increased interest in researching how video games and virtual environments can be used for both learning and assessment of science inquiry skills and practices. Such environments allow students to explore scientific content in a more open-ended context that is designed around actions and choices. In such an environment, students move an avatar around a world, speak to in-game characters, obtain objects, and take those objects to laboratories to run specific tests. While these environments allow for more autonomy and choice, assessing skills in these environments is a more difficult challenge than in closed environments or simulations. In this paper, we present models that can infer two aspects of middle-school students’ inquiry skill, from their interactive behaviors within an assessment in a virtual environment called a “virtual performance assessment” or VPA: 1) whether the student successfully demonstrates the skill of designing controlled experiments within the VPA, and 2) whether a middle-school student can successfully use their inquiry skill to determine the answer to a scientific question with a non-intuitive in-game answer.

Abstract: This paper presents an experimental evaluation of eye gaze data as a source for modeling user’s learning in Interactive Simulations (IS). We com-pare the performance of classifier user models trained only on gaze data vs. models trained only on interface actions vs. models trained on the combination of these two sources of user interaction data. Our long-term goal is to build user models that can trigger adaptive support for students who do not learn well with ISs, caused by the often unstructured and open-ended nature of these environments. The test-bed for our work is the CSP applet, an IS for Constraint Satis-faction Problems (CSP). Our findings show that including gaze data as an additional source of information to the CSP applet’s user model significantly improves model accuracy compared to using interface actions or gaze data alone.

Abstract: Student engagement and motivation during learning activities is tied to better learning behaviors and outcomes and has prompted the development of learner-guided environments. These systems attempt to personalize learning by allowing students to select their own tasks and activities. However, recent evidence suggests that not all students are equally capable of guiding their own learning. Some students are highly self-regulated learners and are able to select learning goals, identify appropriate tasks and activities to achieve these goals and monitor their progress resulting in improved learning and motivational benefits over traditional learning tasks. Students who lack these skills are markedly less successful in self-guided learning environments and require additional scaffolding to be able to navigate them successfully. Prior work has examined these phenomena within the learner-guided environment, CRYSTAL ISLAND, and identified the need for early prediction of students’ self-regulated learning abilities. This work builds upon these findings and presents a dynamic Bayesian approach that significantly improves the classification accuracy of student self-regulated learning skills.

Abstract: A new generation of curation services provides users with a set of tools to manually curate and manage topical collections of content. However, given curation is ultimately a manual effort, it still requires significant effort on the part of the curator both in terms of collecting and managing content. We are interested in providing additional assistance to users in their curation tasks, in particular when it comes to efficiently adding content to their collection, and examine recommender systems in an effort to automate this task. We examine a number of recommendation strategies using live-user data from the popular Scoop.it curation service.

Abstract: It is well-known that for a variety of search tasks involving queries more relevant results can be presented if they are personalized according to a user’s interests and search behavior. This can be achieved with user-dependent, rich web search queries profiles. These are typically built as part of a specific search personalization task so that it is unclear which characteristics of queries are most effective for modeling the user-query relationship in general. In this paper, we explore various approaches for explicitly modeling this user-query relationship independently of other search components. Our models employ generative models in layers in a prediction task. The results show that the best signals for modeling the user-query relationship come from the given query’s terms and entities together with information from related entities and terms, yielding a relative improvement of up to 24.5% in MRR and Success over the baseline methods.

Abstract: This thesis will investigate methods to increase the utility of news article recommendation services. Access to different news providers allows us to consider cross-domain user preferences. We deal with recommender systems with continuously changing item collections. We will be able to observe user feedback from a real-world recommendation system operating on different domains. We will evaluate how results from existing data sets correspond to actual user reactions.

Motivations: Recommender Systems (RS) are information tools designed to suggest items that suit users needs and preferences. They can also support users to browse a product catalogue and better understand and elicit their preferences. These activities are managed by Conversational RSs, which over a series of user-system interactions acquire and revise user preferences by observing the user reaction to proposed options. In this research we focus on the suggestion of queries. We address the problem of helping a user to revise queries for searching in a product catalog with a conversational approach. We want to provide query suggestions that are likely to retrieve products with the largest utility increase, compared to the products retrieved in the previous interaction step. Suggesting query revision is a difficult task given that we do not know the user utility and we do not want to explicitly ask about it. Actually, by observing the query revision selected by the user we can infer some constraints on the user utility function and use this information in order to provide good query revisions. For example, suppose a user queries a product catalogue by issuing a query, such as “I want an hotel with AC and parking”. The system, rather than recommending immediately the products that satisfy this query, assumes that the user may have also other needs and suggests some query revisions. A new query may add an additional feature to the current query, e.g., “are you interested also in sauna?”. Products with more features, if available, will surely increase the user utility. But not all features are equally important for the user. In this approach products are described by their features. The user utility function, also called as user profile, is modeled as a vector of weights which represents the importance the user assigns to each feature. The main challenge is to unveil the weights values which describe the user profile (i.e., the user preferences) by leveraging the information derived from knowing what revised query the user selects among those suggested by the system. This approach was introduced by Ricci 2007 but it has some limitations that we want to address: 1) the high computational cost of computing the best query revisions; 2) the large number of suggestions at each interaction step; 3) it considers only Boolean features; 4) the imposibility to interact with users not having stable preferences.

Abstract: Cross-domain recommender systems suggest items in a target domain by exploiting user preferences and/or domain knowledge available in a source domain. In this thesis we aim to develop a framework for cross-domain recommendation capable of mining heterogeneous sources of information such as semantically annotated data, user generated contents, and contextual signals. For this purpose, we investigate a number of approaches to extract, process, and integrate knowledge for linking distinct domains, and various models that exploit such knowledge for making effective recommendations across domains.

Abstract: In recent years a variety of ontology-based recommender systems, which make use of a domain ontology to characterize the user model, have shown to be very effective. There are however some open issues with this approach, such as: 1) the creation of an ontology is an expensive process; 2) the ontology seldom takes into account the perspectives of target user communities; 3) different groups of users may have different domain conceptualizations; 4) the ontology is usually static and not able to learn automatically new semantic relationships or properties. To address these points, I propose an approach to automatically build multiple personal ontology views (POVs) from user feedbacks, tailored to specific user groups and exploited for recommendation purpose via spreading activation techniques.

Abstract: The Brain-Computer interface (BCI) advancements made possible the use of techniques to recognize emotional aspects from the electroencephalographic signal (EEG). In this work I focus on the implementation of a BCI-based application, able to mine relevant information about user’s emotion from his/her EEG signal and to adapt to it. To this aim a highly low cost and wearable device is employed, so as, a natural interaction is allowed.

Abstract: Emotions and learning are closely related. In the PhD research presented in this paper, that relation has to be taken advantage of. With this aim, within the framework of affective computing, the main goal proposed is modeling learner’s affective state in order to support adaptive features and provide an inclusive personalized e-learning experience. At the first stage of this research, emotion detection is the principal issue to cope with. A multimodal approach has been proposed, so gathering data from diverse sources to feed data mining systems able to supply emotional information is being the current ongoing work. On the next stages, the results of these data mining systems will be used to enhance learner models and based on these, offer a better e-learning experience to improve learner’s results.

Abstract: In our research, we focus on improving the user model by using novel sources of user feedback – tabbed browsing behavior of the users (also called parallel browsing). The tabbing is nowadays established as the more accurate description of browsing activities than the previous linear representation. Users take advantage of multiple tabs in various scenarios, by which they express dif-ferent relations and preferences to hypermedia being visited in such tabs. The aimed contribution is to include this behavior into the user model, so improving accuracy of modeled user’s characteristics and thus improving personalization.

Abstract: Online collections of museums are often hard to access, because the artworks lack appropriate annotations. We develop a framework that supports niches of experts in the crowd in adding annotation of high quality. This thesis focuses on search strategies that match experts with artworks to annotate. Our approach uses explicit semantics for modeling the relations between the properties of the collection items, content-based filtering aimed at diversification, and trust-aware ranking of the results.

Abstract: Secondary user information sharing and usage for purposes other than what it was primarily collected for has become an increasing trend, especially, as we witness a surge in the volume of data collected from and about users online. Although, allowing secondary sharing and usage of data in new and innovative ways is beneficial to the user and the society at large, it also poses the privacy risks of sharing and using personal information for unintended purposes. This paper discusses my PhD thesis towards creating a privacy framework for secondary user information sharing. The aim is to develop an infrastructure that enables sharing of user information across applications and services for beneficial purposes, while balancing it with protecting the user against the potential privacy risks. This paper discusses current work and open challenges.

Abstract: We present a so-called supervised educational recommendation framework in this paper aiming to recommend those programming tasks for a student which improve his skills and performance. The main issue of this approach is an appropriate student model w.r.t. his skills and other implicit factors. The student model can be derived from the solutions provided by the student and the teacher’s (textual as well as numerical) evaluation of these solutions.

Abstract: Social software have been successful in gathering a large number of users in the industrialized world. An opportunity exists in utilizing social software to enhance the quality of life of ourselves and those important to us. In our research we focus on elementary school children as they begin to discover their surrounding areas, and become more involved in interaction with their peers. We explore the possibility of providing socio-geographical support by creating a system of electronic partners (or ePartners), which are intelligent agents that function as teammates to their human users. Since social contexts and familial situations can vary, it is crucial that ePartners are capable of providing personalized support. We aim to achieve that by providing a rich specification language, allowing users to enter their social requirements into the
ePartner as norms.

Abstract: Spellcheckers/predictors can help people in writing more efficiently. It is a well-known fact, for example, that spellcheckers/predictors can ease writing for people with dyslexia. However, most of the spellcheckers assume that wrong words contain just few errors (the literature claims that 80% to 95% of spelling errors contain one error), in terms of the four classical edit operation (i.e., addition, deletion, transposition, substitution), and that errors are isolated (i.e., each error involves just one word). In addition, since standard spellcheckers do not use context, they are not able to correct real-word errors. Finally, they usually are not predictors. This feature is very useful for people with dyslexia, as it allows them to type less characters. The aim of my research is to address the aspect of adaptation and personalization to the individual behavior for the model and the user interface of spellchecker/predictor, considering people with dyslexia. Specifically, we designed and trained a model that takes into account the typical errors (even real-word errors) made by people with dyslexia and the context for spellchecking and prediction, and the experiments to carry out for evaluating its performance. In addition, we formalized the parameters for making the interface adaptive, so that the user interaction with the system is light. In the next months, we will finish the development of the adaptive user interface. Then we will conduct experimental studies for testing the system. From a broader perspective, we try to generalize the system to other user types.