User Modeling, Adaptation, and Personalization

A growing body of work in psychology and the neurosciences has documented the functional, often adaptive role of emotions in human behavior. This has led to a significant growth in research on computationalmodels of human emotional processes, fueled both by their basic research potential as well as the promise that the function of emotion in human behavior can be exploited in a range of applications. Computational models transform theory construction by providing a framework for studying emotion processes that augments what is feasible in more traditional laboratory settings. Modern research in the psychological processes and neural underpinnings of emotion is also transforming the science of computation. In particular, findings on the role that emotions play in human behavior have motivated artificial intelligence and robotics research to explore whether modeling emotion processes can lead to more intelligent, flexible and capable systems. Further, as research has revealed the deep role that emotion and its expression play in human social interaction, researchers have proposed that more effective human computer interaction can be realized if the interaction is mediated both by a model of the user’s emotional state as well as by the expression of emotions.

AdHeat is our newly developed social ad model considering user influence in addition to relevance for matching ads. Traditionally, ad placement employs the relevance model. Such a model matches ads with Web page content, user interests, or both. We have observed, however, on social networks that the relevance model suffers from two shortcomings. First, influential users (users who contribute opinions) seldom click ads that are highly relevant to their expertise. Second, because influential users’ contents and activities are attractive to other users, hint words summarizing their expertise and activities may be widely preferred. Therefore, we propose AdHeat, which diffuses hint words of influential users to others and then matches ads for each user with aggregated hints. Our experimental results on a large-scale social network show that AdHeat outperforms the relevance model on CTR (click through rate) by significant margins. In this talk, the algorithms employed by AdHeat and solutions to address scalability issues are presented.

Adaptive visualization can present user-adaptive information in such a way as to help users to analyze complicated information spaces easily and intuitively. We presented an approach called Adaptive VIBE, which extended the traditional reference point-based visualization algorithm, so that it could adaptively visualize documents of interest. The adaptive visualization was implemented by separating the effects of user models and queries within the document space and we were able to show the potential of the proposed idea. However, adaptive visualization still remained in the simple

bag-of-words

realm. The keywords used to construct the user models were not effective enough to express the concepts that need to be included in the user models. In this study, we tried to improve the old-fashioned keyword-only user models by adopting more concept-rich named-entities. The evaluation results show the strengths and shortcomings of using named-entities as conceptual elements for visual user models and the potential to improve the effectiveness of personalized information access systems.

While browsing the Web, providing profile information in social networking services, or tagging pictures, users leave a plethora of traces. In this paper, we analyze the nature of these traces. We investigate how user data is distributed across different Web systems, and examine ways to aggregate user profile information. Our analyses focus on both explicitly provided profile information (name, homepage, etc.) and activity data (tags assigned to bookmarks or images). The experiments reveal significant benefits of interweaving profile information: more complete profiles, advanced FOAF/vCard profile generation, disclosure of new facets about users, higher level of self-information induced by the profiles, and higher precision for predicting tag-based profiles to solve the cold start problem.

Search engines are key components in the online world and the choice of search engine is an important determinant of the user experience. In this work we seek to model user behaviors and determine key variables that affect search engine usage. In particular, we study the engine usage behavior of more than ten thousand users over a period of six months and use machine learning techniques to identify key trends in the usage of search engines and their relationship with user satisfaction. We also explore methods to determine indicators that are predictive of user trends and show that accurate predictive user models of search engine usage can be developed. Our findings have implications for users as well as search engine designers and marketers seeking to better understand and retain their users.

Today most of existing personalization systems (e.g. content recommenders, or targeted ad) focus on individual users and ignore the social situation in which the services are consumed. However, many human activities are social and involve several individuals whose tastes and expectations must be taken into account by the service providers. When a group profile is not available, different profile aggregation strategies can be applied to recommend adequate content and services to a group of users based on their individual profiles. In this paper, we consider an approach intended to determine the factors that influence the choice of an aggregation strategy. We present a preliminary evaluation made on a real large-scale dataset of TV viewings, showing how group interests can be predicted by combining individual user profiles through an appropriate strategy. The conducted experiments compare the group profiles obtained by aggregating individual user profiles according to various strategies to the “reference” group profile obtained by directly analyzing group consumptions.

Intelligent tutoring systems that utilize Bayesian Knowledge Tracing have achieved the ability to accurately predict student performance not only within the intelligent tutoring system, but on paper post-tests outside of the system. Recent work has suggested that contextual estimation of student guessing and slipping leads to better prediction within the tutoring software (Baker, Corbett, & Aleven, 2008a, 2008b). However, it is not yet clear whether this new variant on knowledge tracing is effective at predicting the latent student knowledge that leads to successful post-test performance. In this paper, we compare the Contextual-Guess-and-Slip variant on Bayesian Knowledge Tracing to classical four-parameter Bayesian Knowledge Tracing and the Individual Difference Weights variant of Bayesian Knowledge Tracing (Corbett & Anderson, 1995), investigating how well each model variant predicts post-test performance. We also test other ways to utilize contextual estimation of slipping within the tutor in post-test prediction, and discuss hypotheses for why slipping during tutor use is a significant predictor of post-test performance, even after Bayesian Knowledge Tracing estimates are controlled for.

This research paper presents the positive effect of incorporating individuals’ working memory (WM) span as a personalization factor in terms of improving users’ academic performance in the context of adaptive educational hypermedia. The psychological construct of WM is robustly related to information processing and learning, while there is a wide differentiation of WM span among individuals. Hence, in an effort to examine the role of cognitive and affective factors in adaptive hypermedia along with psychometric user profiling considerations, WM has a central role in the authors’ effort to develop a user information processing model. Encouraged by previous findings, a larger scale study has been conducted with the participation of 230 university students in order to elucidate if it is possible through personalization to increase the performance of learners with lower levels of WM span. According to the results, users with low WM performed better in the personalized condition, which involved segmentation of the web content and aesthetical annotation, while users with medium/high WM span were slightly negatively affected by the same techniques. Therefore, it can by supported it is possible to specifically address the problem of low WM span with significant results.

Adaptive scaffolding has been proposed as an efficient means for supporting self-directed learning both in educational as well as in adaptive learning systems research. However, the effects of adaptation on self-directed learning and the differential contributions of different adaptation models have not been systematically examined. In this paper, we examine whether personalized scaffolding in the learning process improves learning. We conducted a controlled lab study in which 29 students had to solve several tasks and learn with the help of an adaptive learning system in a within-subjects control condition design. In the learning process, participants obtained recommendations for learning goals from the system in three conditions: fixed scaffolding where learning goals were generated from the domain model, personalized scaffolding where these recommendations were ranked according to the user model, and random suggestions of learning goals (control condition). Students in the two experimental conditions clearly outperformed students in the control condition and felt better supported by the system. Additionally, students who received personalized scaffolding selected fewer learning goals than participants from the other groups.

We develop a multimedia instruction system for the inheritance of skills. This system identifies the difficult segments of video by analyzing user behavior. Difficulties may be inferred by the learner’s requiring more time to fully process a portion of video; they may replay or pause the video during the course of a segment, or play it at a slow speed. These difficult video segments are subsequently assumed to require the addition of expert, instructor annotations, in order to enable learning. We propose a time-dependent annotation mechanism, employing a level of detail (LoD) approach. This annotation is superimposed upon the video, based on the user’s selected speed of playback. The LoD, which reflects the difficulty of the training material, is used to adapt whether to display the annotation to the user. We present the results of an experiment that describes the relationship between the difficulty of material and the LoDs.

Hybridising user models can improve predictive accuracy. However, research on linearly combining predictive user models (e.g., used in recommender systems) has often made the implicit assumption that the individual models perform uniformly across the user and item space, using static model weights when computing a weighted average of the predictions of the individual models. This paper proposes a weighting scheme which combines user- and item-specific weight vectors to compute user- and item-aware model weights. The proposed hybridisation approach adaptively estimates online the model parameters that are specific to a target user as information about this user becomes available. Hence, it is particularly well-suited for domains where little or no information regarding the target user’s preferences or interests is available at the time of offline model training. The proposed weighting scheme is evaluated by applying it to a real-world scenario from the museum domain. Our results show that in our domain, our hybridisation approach attains a higher predictive accuracy than the individual component models. Additionally, our approach outperforms a non-adaptive hybrid model that uses static model weights.

The increasing trend towards powerful mobile phones opens many possibilities for valuable personalised services to be available on the phone. Client-side personalisation for these services has important benefits when connectivity to the cloud is restricted or unavailable. The user may also find it desirable when they prefer that their user model be kept only on their phone and under their own control, rather than under the control of the cloud-based service provider. This paper describes PersonisJ, a user modelling framework that can support client-side personalisation on the Android phone platform. We discuss the particular challenges in creating a user modelling framework for this platform. We have evaluated PersonisJ at two levels: we have created a demonstrator application that delivers a personalised museum tour based on client-side personalisation; we also report on evaluations of its scalability. Contributions of this paper are the description of the architecture, the implementation, and the evaluation of a user modelling framework for client-side personalisation on mobile phones.

In this paper, we present the results of a two-month field study of fifteen people using a software tool designed to model changes in a user’s availability. The software uses status update messages, as well as sensors, to detect changes in context. When changes are identified using the Kullback-Leibler Divergence metric, users are prompted to broadcast their current context to their social networks. The user interface method by which the alert is delivered is evaluated in order to minimize the impact on the user’s workflow. By carefully coupling both algorithms and user interfaces, interruptions made by the software tool can be made valuable to the user.

Recent progress has been made by using sensors with Intelligent Tutoring Systems in classrooms in order to predict the affective state of students users. If tutors are able to interpret sensor data with new students based on past experience, rather than having to be individually trained, then this will enable tutor developers to evaluate various methods of adapting to each student’s affective state using consistent predictions. In the past, our classifiers have predicted student emotions with an accuracy between 78% and 87%. However, it is still unclear which sensors are best, and the educational technology community needs to know this to develop better than baseline classifiers, e.g. ones that use only frequency of emotional occurrence to predict affective state. This paper suggests a method to clarify classifier ranking for the purpose of affective models. The method begins with a careful collection of a training and testing set, each from a separate population, and concludes with a non-parametric ranking of the trained classifiers on the testing set. We illustrate this method with classifiers trained on data collected in the Fall of 2008 and tested on data collected in the Spring of 2009. Our results show that the classifiers for some affective states are significantly better than the baseline model; a validation analysis showed that some but not all classifier rankings generalize to new settings. Overall, our analysis suggests that though there is some benefit gained from simple linear classifiers, more advanced methods or better features may be needed for better classification performance.

Effective pedagogical strategies are important for e-learning environments. While it is assumed that an effective learning environment should craft and adapt its actions to the user’s needs, it is often not clear how to do so. In this paper, we used a Natural Language Tutoring System named Cordillera and applied Reinforcement Learning (RL) to induce pedagogical strategies directly from pre-existing human user interaction corpora. 50 features were explored to model the learning context. Of these features, domain-oriented and system performance features were the most influential while user performance and background features were rarely selected. The induced pedagogical strategies were then evaluated on real users and results were compared with pre-existing human user interaction corpora. Overall, our results show that RL is a feasible approach to induce effective, adaptive pedagogical strategies by using a relatively small training corpus. Moreover, we believe that our approach can be used to develop other adaptive and personalized learning environments.

A long standing challenge for intelligent tutoring system (ITS) designers and educators alike is how to encourage students to take pleasure and interest in learning activities. In this paper, we present findings from a user study involving students interacting with an ITS, focusing on when students express excitement, what we dub “yes!” moments. These findings include an empirically-based user model that relies on both interaction and physiological sensor features to predict “yes!” events; here we describe this model, its validation, and initial indicators of its importance for understanding and fostering student interest.

The overload of the information available on the web, held with the diversity of the user information needs and the ambiguity of their queries have led the researchers to develop personalized search tools that return only documents that meet the user profile representing his main interests and needs. We present in this paper a personalized document ranking model based on an extended graph-based distance measure that exploits a semantic user profile derived from a predefined web ontology (ODP). The measure is based on combining Minimum Common Supergraph (

MCS

) and Maximum Common Subgraph (

mcs

) between graphs representing respectively the document and the user profile. We extend this measure in order to take into account a semantic recovery between the document and the user profile through common concepts and cross links connecting the two graphs. Results show the effectiveness of our personalized graph-based ranking model compared to Yahoo search results.

A user’s informational need and preferences can be modeled by criteria, which in turn can be used to prioritize candidate results and produce a ranked list. We examine the use of such a criteria-based user model separately in two representative recommendation tasks: news article recommendations and product recommendations. We ask the following: are there nonlinear

interactions

among the criteria; and should the models be

personalized

? We assume that that user ratings on each criterion are available, and use machine learning to infer a user model that combines these multiple ratings into a single overall rating. We found that the ratings of different criteria have a nonlinear interaction in some cases, for example, article novelty and subject relevance often interact. We also found that these interactions vary from user to user.

Sentiment analysis deals with inferring people’s sentiments and opinions from texts. An important aspect of sentiment analysis is polarity classification, which consists of inferring a document’s polarity – the overall sentiment conveyed by the text – in the form of a numerical rating. In contrast to existing approaches to polarity classification, we propose to take the authors of the documents into account. Specifically, we present a nearest-neighbour collaborative approach that utilises novel models of user similarity. Our evaluation shows that our approach improves on state-of-the-art performance, and yields insights regarding datasets for which such an improvement is achievable.

User models are generally created to personalise information or share user experiences among like-minded individuals. An individual’s characteristics are compared to those of some canonical user type, and the user included in various user groups accordingly. Those user groups might be defined according to academic ability or recreational interests, but the aim is to include the user in relevant groups where appropriate. The user model described here operates on the principle of exclusion, not inclusion, and its purpose is to detect atypical behaviour, seeing if a user falls outside a category, rather than inside one. That is, it performs anomaly detection against either an individual user model or a typical user model. Such a principle can be usefully applied in many ways, such as early detection of illness, or discovering students with learning issues. In this paper, we apply the anomaly detection principle to the detection of intruders on a computer system masquerading as real users, by comparing the behaviour of the intruder with the expected behaviour of the user as characterised by their user model. This behaviour is captured in characteristics such as typing habits, Web page usage and application usage. An experimental intrusion detection system (IDS) was built with user models reflecting these characteristics, and it was found that comparison with a small number of key characteristics from a user model can very quickly detect anomalies and thus identify an intruder.

User models are a key component for user-adaptive systems. They represent information about users such as interests, expertise, goals, traits, etc. This information is used to achieve various adaptation effects, e.g., recommending relevant documents or products. To ensure acceptance by users, these models need to be scrutable, i.e., users must be able to view and alter them to understand and if necessary correct the assumptions the system makes about the user. However, in most existing systems, this goal is not met. In this paper, we introduce IntrospectiveViews, an interface that enables the user to view and edit her user model. Furthermore, we present the results of a formative evaluation that show the importance users give in general to different aspects of scrutable user models and also substantiate our claim that IntrospectiveViews is an appropriate realization of an interface to such models.

The effectiveness of personalized support provided to virtual communities depends on what we know about a particular community and in which areas the community may need support. Following organizational psychology theories, we have developed algorithms to automatically detect patterns of knowledge sharing in a closely-knit virtual community, focusing on transactive memory, shared mental models, and cognitive centrality. The automatic detection of problematic areas enables taking decisions about notifications targeted at different community members but aiming at improving the functioning of the community as a whole. The paper presents graph-based algorithms for detecting community knowledge sharing patterns, and illustrates, based on a study with an existing community, how these patterns can be used for community-tailored support.

When a quantitative student model is constructed, one of the first tasks to perform is to identify the domain concepts assessed. In general, this task is easily done by the domain experts. In addition, the model may include some misconceptions which are also identified by these experts. Identifying these misconceptions is a difficult task, however, and one which requires considerable previous experience with the students. In fact, sometimes it is difficult to relate these misconceptions to the elements in the knowledge diagnostic system which feeds the student model. In this paper we present a data-driven technique which aims to help elicit the domain misconceptions. It also aims to relate these misconceptions with the assessment activities (e.g. exercises, problems or test questions), which assess the subject in question.

The field of intelligent tutoring systems has been using the well known knowledge tracing model, popularized by Corbett and Anderson (1995), to track student knowledge for over a decade. Surprisingly, models currently in use do not allow for individual learning rates nor individualized estimates of student initial knowledge. Corbett and Anderson, in their original articles, were interested in trying to add individualization to their model which they accomplished but with mixed results. Since their original work, the field has not made significant progress towards individualization of knowledge tracing models in fitting data. In this work, we introduce an elegant way of formulating the individualization problem entirely within a Bayesian networks framework that fits individualized as well as skill specific parameters simultaneously, in a single step. With this new individualization technique we are able to show a reliable improvement in prediction of real world data by individualizing the initial knowledge parameter. We explore three difference strategies for setting the initial individualized knowledge parameters and report that the best strategy is one in which information from multiple skills is used to inform each student’s prior. Using this strategy we achieved lower prediction error in 33 of the 42 problem sets evaluated. The implication of this work is the ability to enhance existing intelligent tutoring systems to more accurately estimate when a student has reached mastery of a skill. Adaptation of instruction based on individualized knowledge and learning speed is discussed as well as open research questions facing those that wish to exploit student and skill information in their user models.

Recently, detectors of gaming the system have been developed for several intelligent tutoring systems where the problem-solving process is reified, and gaming consists of systematic guessing and help abuse. Constraint-based tutors differ from the tutors where gaming detectors have previously been developed on several dimensions: in particular, higher-level answers are assessed according to a larger number of finer-grained constraints, and feedback is split into levels rather than an entire help sequence being available at any time. Correspondingly, help abuse behaviors differ, including behaviors such as rapidly repeating the same answer or blank answers to elicit answers. We use text replay labeling in combination with educational data mining methods to create a gaming detector for SQL-Tutor, a popular constraint-based tutor. This detector assesses gaming at the level of multiple-submission sequences and is accurate both at identifying gaming within submission sequences and at identifying how much each student games the system. It achieves only limited success, however, at distinguishing different types of gaming behavior from each other.

In this paper, we focus on the challenge that users face in processing messages on the web posted in participatory media settings, such as blogs. It is desirable to recommend to users a restricted set of messages that may be most valuable to them. Credibility of a message is an important criteria to judge its value. In our approach, theories developed in sociology, political science and information science are used to design a model for evaluating the credibility of messages that is user-specific and that is sensitive to the social network in which the user resides. To recommend new messages to users, we employ Bayesian learning, built on past user behaviour, integrating new concepts of context and completeness of messages inspired from the strength of weak ties hypothesis, from social network theory. We are able to demonstrate that our method is effective in providing the most credible messages to users and significantly enhances the performance of collaborative filtering recommendation, through a user study on the digg.com dataset.

Our previous research indicates that using personality quizzes is a viable and promising way to build user profiles to recommend entertainment products. Based on these findings, our current research further investigates the feasibility of using personality quizzes to build user profiles not only for an active user but also his or her friends. We first propose a general method that infers users’ music preferences in terms of their personalities. Our in-depth user studies show that while active users perceive the recommended items to be more accurate for their friends, they enjoy more using personality quiz based recommenders for finding items for themselves. Additionally, we explore if domain knowledge has an influence on users’ perception of the system. We found that novice users, who are less knowledgeable about music, generally appreciated more personality-based recommenders. Finally, we propose some design issues for recommender systems using personality quizzes.

User modeling is a complex task, and many user modeling techniques are proposed in the existing literature, but the way these models are presented is not homogeneous, the domain is fragmented and these models are not directly comparable. Thus there is a need for a unified view of the whole user modeling domain and of the applicability of the models to specific applications, contexts or according to specific requirements, type of data, availability of data, etc. A common question companies may ask when they want to build and exploit a user model in order to implement different kinds of personalization or adaptive systems is: “Given my specific requirements, which user modeling technique can be used?”. No obvious answer can be given to this question. This article aims to propose a topic map of user modeling in connection with input data, data types, accessibility, approach, specific requirements and users’ data acquisition methods. This schema/topic map is aimed to help practitioners and researchers as well to answer the above mentioned question. Furthermore the article provides two concrete scenarios in the area of recommender systems and shows how the topic map may be used for these scenarios and real world applications.

myCOMAND case study explores the vision of an interactive user interface (UI) in the vehicle providing access to a large variety of information items aggregated from Web services. It was created for gaining insights into applicability of personalization and recommendation approaches for the visual ranking and grouping of items, composed as interactive UI layout components (e.g. carousels, lists). Quick access to preferred and important items can support less distracting interaction with a large web-based content collections and smaller screen size. Content gets aggregated on the server and then synchronized to an onboard module. Ranking for each data item is annotated based on a user profiles with a fuzzy preferences and a shared taxonomy on content categories. Preference values are implicitly learned from user interaction, but can be set explicitly by the user too. A circular UI component for browsing Internet radio stations is described, which dynamically groups items into categories during scrolling. Items are ranked according to the user’s preferences and item novelty. A visual overview mode helps to quickly review the structure of large content collections.

Telecommunication applications based on user modeling focus on extracting customer behavior and preferences from the information implicitly included in Call Detail Record (CDR) datasets. Even though there are many different application areas (fraud detection, viral and targeted marketing, churn prediction, etc.) they all share a common data source (CDRs) and a common set of features for modeling the user. In this paper we present our experience with different applications areas in generating user models from massive real datasets of both mobile phone and landline subscriber activity. We present the analysis of a dataset containing the traces of 50,000 mobile phone users and 50,000 landline users from the same geographical area for a period of six months and compare the different behaviors when using landlines and mobile phones and the implications that such differences have for each application. Our results indicate that user models for a variety of applications can be generated efficiently and in a homogeneous way using an architecture based on distributed computing and that there are numerous differences between mobile phone and landline users that have relevant practical implications.

The World Wide Web has provided users with the opportunity to access from any computer the largest set of information ever existing. Researchers have analyzed how such users surf the Web, and such analysis has been used to improve existing services (e.g., by means of data mining and personalization techniques) as well as the generation of new ones (e.g., online targeted advertisement). In recent years, a new trend has developed by which users do not need a computer to access the Web. Instead, the low prices of mobile data connections allow them to access it anywhere anytime. Some studies analyze how users access the Web on their handsets, but these studies use only navigation logs from a specific portal. Therefore, very little attention (due to the complexity of obtaining the data) has been given to how users surf the Web (off-portal) from their mobiles and how that information could be used to build user profiles. This paper analyzes full navigation logs of a large set of mobile users in a developed country, providing useful information about the way those users access the Web. Additionally, it explores how navigation logs can be categorized, and thus user’s interest can be modeled, by using online sources of information such as Web directories and social tagging systems.

Recommender systems typically require feedback from the user to learn the user’s taste. This feedback can come in two forms: explicit and implicit. Explicit feedback consists of ratings provided by the user for a number of items, while implicit feedback comes from observing user actions on items. These actions have to be interpreted by the recommender system and translated into a rating. In this paper we propose a method to learn how to translate user actions on items to ratings on these items by correlating user actions with explicit feedback. We do this by associating user actions to rated items and subsequently applying naive Bayesian classification to rate new items with which the user has interacted. We apply and evaluate our method on data from a web-based music service and we show its potential as an addition to explicit rating.

This paper proposes a personalised frequency-based model for predicting a user’s pathway through a physical space, based on non-intrusive observations of users’ previous movements. Specifically, our approach estimates a user’s transition probabilities between discrete locations utilising personalised transition frequency counts, which in turn are estimated from the movements of other similar users. Our evaluation with a real-world dataset from the museum domain shows that our approach performs at least as well as a non-personalised frequency-based baseline, while attaining a higher predictive accuracy than a model based on the spatial layout of the physical museum space.

In web-based adaptive systems, the same rating scales are usually provided to all users for expressing their preferences with respect to various items. It emerged from a user experiment that we recently carried out that different users show different preferences with respect to the rating scales to use in the interface of adaptive systems, given the particular topic they are evaluating. Starting from this finding, we propose to allow users to choose the kind of rating scale they prefer. This approach raises various issues; the most important is that of how an adaptation algorithm can properly deal with values coming from heterogeneous rating scales. We conducted an experiment to investigate how users rate the same object on different rating scales. On the basis of our interpretation of these results, as an example of one possible solution approach, we propose a three-phase normalization process for mapping preferences expressed with different rating scales onto a unique system representation.

Recommender systems, as a type of Web personalized service to support users’ online product searching, have been widely developed in recent years but with primary emphasis on algorithm accuracy. In this paper, we particularly investigate the efficacy of recommender interface designs in affecting users’ decision making strategies through the observation of their eye movements and product selection behavior. One interface design is the standard list interface where all recommended items are listed one by one. Another two are layout variations of organization-based interface where recommendations are grouped into categories. The eye-tracking user evaluation shows that the organization interfaces, especially the one with a quadrant layout, can significantly attract users’ attentions to more items, with the resulting benefit to enhance their objective decision quality.

With the number of people considered to be obese rising across the globe, the role of IT solutions in health management has been receiving increased attention by medical professionals in recent years. This paper focuses on an initial step toward understanding the applicability of recommender techniques in the food and diet domain. By understanding the food preferences and assisting users to plan a healthy and appealing meal, we aim to reduce the effort required of users to change their diet. As an initial feasibility study, we evaluate the performance of collaborative filtering, content-based and hybrid recommender algorithms on a dataset of 43,000 ratings from 512 users. We report on the accuracy and coverage of the algorithms and show that a content-based approach with a simple mechanism that breaks down recipe ratings into ingredient ratings performs best overall.

Older studies have proved that when searching information on the Web, users tend to write short queries, unconsciously trying to minimize the cognitive load. However, as these short queries are very ambiguous, search engines tend to find the most popular meaning – someone who does not know anything about cascading stylesheets might search for a music band called

css

and be very surprised about the results. In this paper we propose a method which can infer additional keywords for a search query by leveraging a social network context and a method to build this network from the stream of user’s activity on the Web.

Building on previous research with an independent open learner model in a range of university courses, this paper investigates features that may influence student choice about whether to use the environment in a particular course. It was found that some features are considered particularly important by students, but other features are less influential in students’ decisions to use an independent open learner model. Recommendations for features to consider promoting uptake of this type of environment are given.

Emotional intelligence is a clear factor in education [1–3], health care [4], and day to day interaction. With the increasing use of computer technology, computers are interacting with more and more individuals. This interaction provides an opportunity to increase knowledge about human emotion for human consumption, well-being, and improved computer adaptation.

This research makes five main contributions. 1) Construct a method for determining a set of sensor features that can be automatically processed to predict human emotional changes in observed people. 2) Identify principles, algorithms, and classifiers that enable computational recognition of human emotion. 3) Apply this method to an intelligent tutoring system instrumented with sensors. 4) Apply and adapt the method to audio and video sensors for a number of applications such as a) detection of psychological disorders, b) detection of emotional changes in health care providers, c) detection of emotional impact of one person on another during video chat, and/or d) detection of emotional impact of one fictional character on another in a motion picture. 5) Integrate emotional detection technologies so that they can be used in more realistic settings.

Recommender systems traditionally rely on numeric ratings to represent user opinions, and thus are limited by the single-dimensional nature of such ratings. Recent years have seen an abundance of user-generated texts available online, and advances in natural language processing allow us to better understand users by analysing the texts they write. Specifically, sentiment analysis enables inference of people’s sentiments and opinions from texts, while authorship attribution investigates authors’ characteristics. We propose to use these techniques to build text-based user models, and incorporate these models into state-of-the-art recommender systems to generate recommendations that are based on a more profound understanding of the users than rating-based recommendations. Our preliminary results suggest that this is a promising direction.

Semantically-enhanced Ubiquitous User Modeling aims at the management of distributed user models and the integration into ontologies to share user information amongst adaptive applications for personalization purposes. To reach this goal, different problems have to be solved. The collection of implicit user information by observing the user behavior on dynamic web applications is important to better understand the user interests and needs. The aggregation of different user models is essential to combine all available user information to one big knowledge repository. Additionally, the Semantic Web offers new possibilities to enhance the knowledge about the user for better personalization.

In this paper we present an approach to user modeling based on the domain model that we generate

automatically

by resource (text) content processing and analysis of associated tags from a social annotation service. User’s interests are modeled by overlaying the domain model – via keywords extracted from resource’s (text) content, and tags assigned by the user or other (similar) users. The user model is derived automatically. We combine content- and tag-based approaches, shifting our approach beyond flat “folksonomical” representation of user interests to involve relationships between both keywords and tags.

In this paper we are proposing a proof-of-concept leveraging the use of 3D virtual worlds in addictive behavior interventions. We propose a model that we call biographic space, which embeds the successive stages that a smoker may go through while attempting to quit smoking including emotionally loaded aspects such as deciding to quit and post cessation withdrawal. The design of this space is informed by storytelling and explores the rich media affordance of virtual environments.

A Knowledge Worker (KW) uses her computer to perform different tasks for which she gathers and uses information from disparate sources such as the Web and e-mail, and creates new information such as calendar events, e-mails, and documents (

resources

). This forms a Task Space (TS): an information space composed of all computer-based resources the KW uses in relation to a task. Furthermore, KWs may switch between multiple tasks, some of which may be suspended and resumed after some time. These effects compound the KW’s ability to organise and visualise an accurate mental model of the individual TSs. We propose a Task-Based User Model (TBUM) that acts as the KW’s mental model for each task by automatically tracking, relating and organising resources associated with that task. The generated TBUM can be used to support complex activities such as task-resumption, searching within a task-context, task sharing and collaboration.

Leveraging interactive systems by integrating adaptivity is considered as an important key to accommodate user diversity and enhance user interaction. A virtual environment is a highly interactive system which involves users performing complex tasks using diverse 3D interaction techniques. Adaptivity has not been investigated thoroughly in the context of virtual environments. This PhD research is concerned with embedding intelligence to enhance user interaction in virtual environments (i.e. providing adaptive personalized 3D interaction techniques).