Proceedings of the 2011 Conference on User Modeling, Adaptation and Personalization
Time and place:
The biennial conference series User Modeling (UM, 1986-2007) and Adaptive Hypermedia and Adaptive Web-Based Systems (AH, 2000-2008) have been merged into the annual conference series User Modeling, Adaptation, and Personalization (UMAP). UMAP is the most important conference for those interested in any aspect of (interaction with) systems that acquire information about a user (or group of users) so as to be able to adapt their behavior to that user or group.
How can micro-blogging activities on Twitter be leveraged for user modeling and personalization? In this paper we investigate this question and introduce a framework for user modeling on Twitter which enriches the semantics of Twitter messages (tweets) and identifies topics and entities (e.g. persons, events, products) mentioned in tweets. We analyze how strategies for constructing hashtag-based, entity-based or topic-based user profiles benefit from semantic enrichment and explore the temporal dynamics of those profiles. We further measure and compare the performance of the user modeling strategies in context of a personalized news recommendation system. Our results reveal how semantic enrichment enhances the variety and quality of the generated user profiles. Further, we see how the different user modeling strategies impact personalization and discover that the consideration of temporal profile patterns can improve recommendation quality.
Social annotation systems enable the organization of online resources with user-defined keywords. The size and complexity of these systems make them excellent platforms for the application of recommender systems, which can provide personalized views of complex information spaces. Many researchers have concentrated on the important problem of tag recommendation. Less attention has been paid to the recommendation of resources in the context of social annotation systems. In this paper, we examine the specific case of tag-based resource recommendation and propose a linear-weighted hybrid for the task. Using six real world datasets, we show that our algorithm is more effective than other more mathematically complex techniques.
As showed in a previous work, different users show different preferences with respect to the rating scales to use for evaluating items in recommender systems. Thus in order to promote users' participation and satisfaction with recommender systems, we propose to allow users to choose the rating scales to use. Thus, recommender systems should be able to deal with ratings coming from heterogeneous scales in order to produce correct recommendations. In this paper we present two user studies that investigate the role of rating scales on user's rating behavior, showing that the rating scales have their own "personality" and mathematical normalization is not enough to cope with mapping among different rating scales.
Over the last decades, there have been a rich variety of approaches towards modeling student knowledge and skill within interactive learning environments. There have recently been several empirical comparisons as to which types of student models are better at predicting future performance, both within and outside of the interactive learning environment. However, these comparisons have produced contradictory results. Within this paper, we examine whether ensemble methods, which integrate multiple models, can produce prediction results comparable to or better than the best of nine student modeling frameworks, taken individually. We ensemble model predictions within a Cognitive Tutor for Genetics, at the level of predicting knowledge action-by-action within the tutor. We evaluate the predictions in terms of future performance within the tutor and on a paper post-test. Within this data set, we do not find evidence that ensembles of models are significantly better. Ensembles of models perform comparably to or slightly better than the best individual models, at predicting future performance within the tutor software. However, the ensembles of models perform marginally significantly worse than the best individual models, at predicting post-test performance.
Student modeling plays an important role in educational research. Many techniques have been developed focusing on accurately estimating student performances. In this paper, using Performance Factors Analysis as our framework, we examine what components of the model enable us to better predict, and consequently better understand, student performance. Using transfer models to predict is very common across different student modeling techniques, as student proficiencies on those required skills are believed, to a large degree, to determine student performance. However, we found that problem difficulty is an even more important predictor than student knowledge of the required skills. In addition, we found that using student proficiencies across all skills works better than just using those skills thought relevant by the transfer model. We tested our proposed models with two transfer models of fine- and coarse-grain sizes; the results suggest that the improvement is not simply an illusion due to possible mistakes in associating skills with problems.
The main source of information in most adaptive hypermedia systems are server monitored events such as page visits and link selections. One drawback of this approach is that pages are treated as "monolithic" entities, since the system cannot determine what portions may have drawn the user's attention. Departing from this model, the work described here demonstrates that client-side monitoring and interpretation of users' interactive behavior (such as mouse moves, clicks and scrolling) allows for detailed and significantly accurate predictions on what sections of a page have been looked at. More specifically, this paper provides a detailed description of an algorithm developed to predict which paragraphs of text in a hypertext document have been read, and to which extent. It also describes the user study, involving eye-tracking for baseline comparison, that served as the basis for the algorithm.
Competence management systems are increasingly based on ontologies representing competences within a certain domain. Most of these systems represent a user's competence profile by means of an ontological structure. Such semantic competence profiles, often structured as a hierarchy of competences, are difficult to navigate for self-assessment purposes. The more competences a user profile holds, the more challenging the comprehensive presentation of profile data is. In this paper, we present an integrated user interface that supports users during competence self-assessment and facilitates a clear presentation of their semantic competence profiles. For evaluation, we conducted a usability study with 19 students at university. The results show that users were mostly satisfied with the usability of the interface that also represents a promising approach for efficient competence self-assessment.
This paper explores a social extension of open student modeling that we call open social student modeling. We present a specific implementation of this approach that uses parallel IntrospectiveViews to visualize models representing student progress with QuizJET parameterized self-assessment questions for Java programming. The interface allows visualizing not only the student's own model, but also displaying parallel views on the models of their peers and the cumulative model of the entire class or group. The system was evaluated in a semester-long classroom study. While the use of the system was non-mandatory, the parallel IntrospectiveViews interface caused an increase in all of the usage parameters in comparison to a regular portal-based access, which allowed the student to achieve a higher success rate in answering the questions. The collected data offer some evidence that a combination of traditional personalized guidance with social guidance was more effective than personalized guidance alone.
Context-aware music recommender systems are capable to suggest music items taking into consideration contextual conditions, such as the user mood or location, that may influence the user preferences at a particular moment. In this paper we consider a particular kind of context aware recommendation task -- selecting music content that fits a place of interest (POI). To address this problem we have used emotional tags attached by a users' population to both music and POIs. Moreover, we have considered a set of similarity metrics for tagged resources to establish a match between music tracks and POIs. In order to test our hypothesis, i.e., that the users will reckon that a music track suits a POI when this track is selected by our approach, we have designed a live user experiment where subjects are repeatedly presented with POIs and a selection of music tracks, some of them matching the presented POI and some not. The results of the experiment show that there is a strong overlap between the users' selections and the best matching music that is recommended by the system for a POI.
In recent years, social media services with social tagging have become tremendously popular. Because users are no longer mere consumers of content, social Web users have been overwhelmed by the huge numbers of social content available. For tailoring search results, in this paper, we look into the potential of social tagging in social media services. By leveraging collaborative filtering, we propose a new search model to enhance not only retrieval accuracy but also retrieval coverage. Our approach first computes latent preferences of users on tags from other similar users, as well as latent annotations of tags for items from other similar items. We then apply the latency of tags to a tag-based personalized ranking depending on individual users. Experimental results demonstrate the feasibility of our method for personalized searches in social media services.
There are many contexts where it would be helpful to model the collaboration of a group. In learning settings, this is important for classroom teachers and for students learning collaboration skills. Our approach exploits the digital and audio footprints of the users' actions at collocated settings to automatically build a model of symmetry of activity. This paper describes our theoretical model of collaborative learning and how we implemented it. We use the Gini coefficient as a statistical indicator of symmetry of activity, which is itself an important indicator of collaboration. We built this model from a small-scale qualitative study based on concept mapping at an interactive tabletop. We then evaluated the model using a larger scale study based on a corpus of coded data from a multi-display groupware collocated setting. Our key contributions are the model of symmetry of activity as a foundation for modelling collaboration within groups that should have egalitarian participation, the operationalisation of the model and validation of the approach on both a small-scale qualitative study and a larger scale quantitative corpus of data.
Human activity recognition aims to infer the actions of one or more persons from a set of observations captured by sensors. Usually, this is performed by following a fixed length sliding window approach for the features extraction where two parameters have to be fixed: the size of the window and the shift. In this paper we propose a different approach using dynamic windows based on events. Our approach adjusts dynamically the window size and the shift at every step. Using our approach we have generated a model to compare both approaches. Experiments with public datasets show that our method, employing simpler models, is able to accurately recognize the activities, using fewer instances, and obtains better results than the approaches used by the datasets authors.
Question answering communities (QA) are sustained by a handful of experts who provide a large number of high quality answers. Identifying these experts during the first few weeks of their joining the community can be beneficial as it would allow community managers to take steps to develop and retain these potential experts. In this paper, we explore approaches to identify potential experts as early as within the first two weeks of their association with the QA. We look at users' behavior and estimate their motivation and ability to help others. These qualities enable us to build classification and ranking models to identify users who are likely to become experts in the future. Our results indicate that the current experts can be effectively identified from their early behavior. We asked community managers to evaluate the potential experts identified by our algorithm and their analysis revealed that quite a few of these users were already experts or on the path of becoming experts. Our retrospective analysis shows that some of these potential experts had already left the community, highlighting the value of early identification and engagement.
Many models in computer education and assessment take into account difficulty. However, despite the positive results of models that take difficulty in to account, knowledge tracing is still used in its basic form due to its skill level diagnostic abilities that are very useful to teachers. This leads to the research question we address in this work: Can KT be effectively extended to capture item difficulty and improve prediction accuracy? There have been a variety of extensions to KT in recent years. One such extension was Baker's contextual guess and slip model. While this model has shown positive gains over KT in internal validation testing, it has not performed well relative to KT on unseen in-tutor data or post-test data, however, it has proven a valuable model to use alongside other models. The contextual guess and slip model increases the complexity of KT by adding regression steps and feature generation. The added complexity of feature generation across datasets may have hindered the performance of this model. Therefore, one of the aims of our work here is to make the most minimal of modifications to the KT model in order to add item difficulty and keep the modification limited to changing the topology of the model. We analyze datasets from two intelligent tutoring systems with KT and a model we have called KT-IDEM (Item Difficulty Effect Model) and show that substantial performance gains can be achieved with this minor modification that incorporates item difficulty.
Our bodies shape our experience of the world, and our bodies influence what we design. How important are the physical differences between people? Can we model the physiological differences and use the models to adapt and personalize designs, user interfaces and artifacts? Within many disciplines Digital Human Models and Standard Observer Models are widely used and have proven to be very useful for modeling users and simulating humans. In this paper, we create personalized digital human models of perception (Individual Observer Models), particularly focused on how humans see. Individual Observer Models capture how our bodies shape our perceptions. Individual Observer Models are useful for adapting and personalizing user interfaces and artifacts to suit individual users' bodies and perceptions. We introduce and demonstrate an Individual Observer Model of human eyesight, which we use to simulate 3600 biologically valid human eyes. An evaluation of the simulated eyes finds that they see eye charts the same as humans. Also demonstrated is the Individual Observer Model successfully making predictions about how easy or hard it is to see visual information and visual designs. The ability to predict and adapt visual information to maximize how effective it is is an important problem in visual design and analytics.
Most of the approaches for understanding user preferences or taste are based on having explicit feedback from users. However, in many real-life situations we need to rely on implicit feedback. To analyze the relation between implicit and explicit feedback, we conduct a user experiment in the music domain. We find that there is a strong relation between implicit feedback and ratings. We analyze the effect of context variables on the ratings and find that recentness of interaction has a significant effect. We also analyze several user variables. Finally, we propose a simple linear model that relates these variables to the rating we can expect to an item. Such mapping would allow to easily adapt any existing approach that uses explicit feedback to the implicit case and combine both kinds of feedback.
This paper explores ways to address the problem of the high cost problem of poor recommendations in reciprocal recommender systems. These systems recommend one person to another and require that both people like each other for the recommendation to be successful. A notable example, and the focus of our experiments is online dating. In such domains, poor recommendations should be avoided as they cause users to suffer repeated rejection and abandon the site. This paper describes our experiments to create a recommender based on two classes of models: one to predict who each user will like; the other to predict who each user will dislike. We then combine these models to generate recommendations for the user. This work is novel in exploring modelling both people's likes and dislikes and how to combine these to support a reciprocal recommendation, which is important for many domains, including online dating, employment, mentor-mentee matching and help-helper matching. Using a negative and a positive preference model in a combined manner, we improved the success rate of reciprocal recommendations by 18% while, at the same time, reducing the failure rate by 36% for the top-1 recommendations in comparison to using the positive model of preference alone.
Theme parks are important and complex forms of entertainment, with a broad user-base, and with a substantial economic impact. In this paper, we present a case study of an existing theme park, and use this to motivate two research challenges in relation to user-modeling and personalization in this environment: developing recommender systems to support theme park visits, and developing rides that are personalized to the users who take part in them. We then provide an analysis, drawn from a real-world study on an existing ride, which illustrates the efficacy of psychometric profiling and physiological monitoring in relation to these challenges. We conclude by discussing further research work that could be carried out within the theme park, but motivate this research by considering the broader contribution to user-modeling issues that it could make. As such, we present the theme park as a microcosm which is amenable to research, but which is relevant in a much broader setting.
In this paper we focus on an approach to social search, HeyStaks that is designed to integrate with mainstream search engines such as Google, Yahoo and Bing. HeyStaks is motivated by the idea that Web search is an inherently social or collaborative activity. Heystaks users search as normal but benefit from collaboration features, allowing searchers to better organise and share their search experiences. Users can create and share repositories of search knowledge (so-called search staks) in order to benefit from the searches of friends and colleagues. As such search staks are community-based information resources. A key challenge for HeyStaks is predicting which search stak is most relevant to the users current search context and in this paper we focus on this so-called stak recommendation issue by looking at a number of different approaches to profiling and recommending community-search knowledge.
Recommender systems generally face the challenge of making predictions using only the relatively few user ratings available for a given domain. Cross-domain collaborative filtering (CF) aims to alleviate the effects of this data sparseness by transferring knowledge from other domains. We propose a novel algorithm, Tag-induced Cross-Domain Collaborative Filtering (TagCDCF), which exploits user-contributed tags that are common to multiple domains in order to establish the cross-domain links necessary for successful cross-domain CF. TagCDCF extends the state-of-the-art matrix factorization by introducing a constraint involving tag-based similarities between pairs of users and pairs of items across domains. The method requires no common users or items across domains. Using two publicly available CF data sets as different domains, we experimentally demonstrate that TagCDCF substantially outperforms other state-of-the-art single domain CF and cross-domain CF approaches. Additional experiments show that TagCDCF addresses data sparseness and illustrate the influence of the number of tags used by users in both domains.
This paper is intended as guidance for those who are familiar with user modeling field but are less fluent in statistical methods. It addresses potential problems with user model selection and evaluation, that are often clear to expert modelers, but are not obvious for others. These problems are frequently a result of a falsely straightforward application of statistics to user modeling (e.g. over-reliance on model fit metrics). In such cases, absolute trust in arguably shallow model accuracy measures could lead to selecting models that are hard-to-interpret, less meaningful, over-fit, and less generalizable. We offer a list of questions to consider in order to avoid these modeling pitfalls. Each of the listed questions is backed by an illustrative example based on the user modeling approach called Performance Factors Analysis (PFA) .
The success of online social networking systems has revolutionised online sharing and communication, however it has also contributed significantly to the infamous information overload problem. Social Networking systems aggregate network activities into chronologically ordered lists, Network Feeds, as a way of summarising network activity for its users. Unfortunately, these feeds do not take into account the interests of the user viewing them or the relevance of each feed item to the viewer. Consequently individuals often miss out on important updates. This work aims to reduce the burden on users of identifying relevant feed items by exploiting observed user interactions with content and people on the network and facilitates the personalization of network feeds in a manner which promotes relevant activities. We present the results of a large scale live evaluation which shows that personalized feeds are more successful at attracting user attention than non-personalized feeds.
Traditional desktop search paradigm often does not fit mobile contexts. Common mobile devices provide impoverished mechanisms for text entry and small screens are able to offer only a limited set of options, therefore the users are not usually able to specify their needs. On a different note, mobile technologies have become part of the everyday life as shown by the estimate of one billion of mobile broadband subscriptions in 2011. This paper describes an approach to make context-aware mobile interaction available in scenarios where users might be looking for categories of points of interest (POIs), such as cultural events and restaurants, through remote location-based services. Empirical evaluations shows how rich representations of user contexts has the chance to increase the relevance of the retrieved POIs.
Students interacted with an intelligent tutoring system to learn grammatical rules for an artificial language. Six tutoring policies were explored. One, based on a Dynamic Bayes' Network model of skills, was learned from the performance of previous students. Overall, this policy and other intelligent policies outperformed random policies. Some policies allowed students to choose one of three problems to work on, while others presented a single problem at each iteration. The benefit of choice was not apparent in group statistics; however, there was a strong interaction with gender. Overall, women learned less than men, but they learned different amounts in the choice and no choice conditions, whereas men seemed unaffected by choice. We explore reasons for these interactions between gender, choice and learning.
Our goal is to develop methods for non-experts to teach complex behaviors to autonomous agents (such as robots) by accommodating "natural" forms of human teaching. We built a prototype interface allowing humans to teach a simulated robot a complex task using several techniques and report the results of 44 human participants using this interface. We found that teaching styles varied considerably but can be roughly categorized based on the types of interaction, patterns of testing, and general organization of the lessons given by the teacher. Our study contributes to a better understanding of human teaching patterns and makes specific recommendations for future human-robot interaction systems.
Classic adaptive hypermedia systems are able to track a user's knowledge of the subject and use it to evaluate the novelty and difficulty of content encountered by the user. Our goal is to implement this functionality in an open corpus context where a domain model is not available nor is the content indexed with domain concepts. We examine methods for novelty measurement based on automatic text analysis. To compare these methods, we use an evaluation approach based on knowledge encapsulated in the structure of a textbook. Our study shows that a knowledge accumulation method adopted from the domain of intelligent tutoring systems offers a more meaningful novelty measurement than methods adapted from the area of personalized information retrieval.
4-coach Mathematics Active Learning Intelligent Tutoring sYstem (4MALITY) is a web-based intelligent tutoring system for 3rd, 4th, and 5th grade students who are learning math content from the state of Massachusetts (USA) required curriculum framework. The goal of 4MALITY is to personalize help for students by offering them problem-solving strategies authored from multiple points of view. Four virtual coaches (Estella Explainer, Chef Math Bear, How-to Hound, and Visual Vicuna) are designed to capture the character and content of these different problem-solving approaches with language, computation, strategy, and visual hints. A preliminary study was run with 102 students in fourth and fifth grade math classrooms over a period of two months. The results showed that the effect of using 4MALITY produced a statistically significant increase in post-test scores. We explored student performance, help-seeking behavior and meta-cognitive strategies by gender and math ability and report these results.
Uncertainty, diversity and change create endless streams of unexpected new opportunities. To seize those opportunities, new web-based systems are emerging that enforce participative design and empower end-users to take actively part in the creation and maintenance of functionality that fits specific needs and conditions. For example, Yahoo! Pipes is a "participative site" with visual online programming means for defining and readily deploying web-based services that fetch, aggregate and process web feeds. Standard and dedicated engineering tools for developing such web sites are however yet to be invented. This paper describes our software platform for their development by reuse and extension, while meeting the requirements of end-user accessibility, expressivity, interpretability, web compatibility, shareability and traceability as they appear in person-centric areas like Ambient Assisted Living. We allow dynamic and user-driven individualization of functionality by capturing at runtime, and processing complex interaction patterns that involve end-users, their physical environment and software components.
The socioeconomic status of a population or an individual provides an understanding of its access to housing, education, health or basic services like water and electricity. In itself, it is also an indirect indicator of the purchasing power and as such a key element when personalizing the interaction with a customer, especially for marketing campaigns or offers of new products. In this paper we study if the information derived from the aggregated use of cell phone records can be used to identify the socioeconomic levels of a population. We present predictive models constructed with SVMs and Random Forests that use the aggregated behavioral variables of the communication antennas to predict socioeconomic levels. Our results show correct prediction rates of over 80% for an urban population of around 500,000 citizens.
In this paper, we examine a challenge that arises in the application of peer-based tutoring: coping with inappropriate advice from peers. We examine an environment where students are presented with those learning objects predicted to improve their learning (on the basis of the success of previous, like-minded students) but where peers can additionally inject annotations. To avoid presenting annotations that would detract from student learning (e.g. those found confusing by other students) we integrate trust modeling, to detect over time the reputation of the annotation (as voted by previous students) and the reputability of the annotator. We empirically demonstrate, through simulation, that even when the environment is populated with a large number of poor annotations, our algorithm for directing the learning of the students is effective, confirming the value of our proposed approach for student modeling. In addition, the research introduces a valuable integration of trust modeling into educational applications.
We present results from a user study of the Reading Glove version 2.0, a combination wearable and tabletop interactive narrative system. The system was designed to study user perceptions of adaptivity. The system's reasoning engine guides users through the story using three different recommendation modes: random recommendations, story content-based recommendations, and user model based recommendations. We look at the differences in user behaviour and experience across the three recommendation systems, using information from system logs and user surveys and interviews.
Research on Recommender Systems has barely explored the issue of adapting a recommendation strategy to the user's information available at a certain time. In this thesis, we introduce a component that allows building dynamic recommendation strategies, by reformulating the performance prediction problem in the area of Information Retrieval to that of recommender systems. More specifically, we investigate a number of adaptations of the query clarity predictor in order to infer the ambiguity in user and item profiles. The properties of each predictor are empirically studied by, first, checking the correlation of the predictor output with a performance measure, and second, by incorporating a performance predictor into a recommender system to produce a dynamic strategy. Depending on how the predictor is integrated with the system, we explore two different applications: dynamic user neighbour weighting and hybrid recommendation. The performance of such dynamic strategies is examined and compared with that of static ones.
Simulated environments, where learners are involved in simulated situations that resemble actual activities, gain a growing popularity in professional training, and provide powerful experiential learning tools for developing soft skills in ill-defined domains. Adaptation and personalization will play a key role in these environments.
Situation awareness is a perception of the available information, events, resources, and environment within a given time and space. Humans have limited abilities to obtain and maintain situation awareness, as they need to carefully orchestrate the available resources. A failure to maintain situation awareness may lead to serious errors in human behavior. Investigation of the situation awareness of neurosurgeons using cognitive architectures is a new and exciting application of computational user modeling. Accurately modeling of the surgeons' behavior and their mental states while they perform operations using miniature instruments and movements require various implicit measures of the surgeons' behavior. The user modeling community has been searching for such data sources in other domains and have indicated that eye-tracking, as a noninvasive methodology, can be used to enrich the user models and increase their quality. In this research I will 1) investigate what are the constituents of situation awareness during neurosurgery, 2) how eye-tracking methodologies fit to created suitable user models of situation awareness, and 3) how data should be processed, and what features of eye-tracking data work best. We propose to use eye tracking techniques to develop a comprehensive computational model of the surgeons' behavior. The model will be further interpreted, to understand how information, events, and surgeons' actions will impact neurosurgery operations.
Recommender Systems (RSs) generate personalized suggestions to users for items that may be interesting for them. Many RSs use the Collaborative Filtering (CF) technique, where the system gathers some information about the users by eliciting their ratings for items. To do so, the system may actively choose the items to present to the users to rate. This proactive approach is called Active Learning (AL), since the system actively search for relevant data before building any predictive model of the user interests. But, since not all the ratings will improve the accuracy in the same way, finding the best items to query the users for their ratings is challenging. In this work, we address this problem by reviewing some AL techniques and discussing their performance on the base of the experiments we made.
Sound and music online services driven by communities of users are filled with large amounts of user-created content that has to be properly described. In these services, typical sound and music modeling is performed using either content-based or context-based strategies, but no special emphasis is given to the extraction of knowledge from the community. We outline a research plan in the context of Freesound.org and propose ideas about how audio clip sharing sites could adapt and take advantage of particular user communities to improve the descriptions of their content.
This document contains a brief description of my PhD research, with problem definition, contribution to the field of reputation systems and user modeling, and proposed solution. The proposed method and algorithm enable evaluation of contributions in online knowledge-based communities. The innovation in the approach is the use of authority and specifying reputation on the keyword-level.
This paper describes our research lines that focus on modeling and inferring student procedural knowledge in Intelligent Tutoring Systems. Our proposal is to apply Item Response Theory, a well-founded theory for declarative knowledge assessment, to infer procedural knowledge in problem solving environments. Therefore, we treat the problems as tests and the steps of problem solving as options (or choices) in a question. An important feature of our system is that it is not only based on an expert analysis, but also on data-driven techniques so that it can collect the largest amount of students' problem solving strategies as possible.
Personal information management (PIM) is a study on how people handle personal information to support their needs and tasks. In the last decade a lot of studies focused on how people acquire, organize, maintain and retrieve information from their information spaces. Results have led to many research prototypes that tried to either augment present tools or integrate these collections within entirely new designs. However, not much has changed in the present tools, and hierarchies still prevail as the storage foundation. Our research aims at understanding the difference between how people organize their information in various applications and physical space and how they actually think of this information in relation to tasks they have to accomplish. We carried out a preliminary study and are currently finishing another study which both show that there is a difference on how information is organized in formal structures on computers and physical spaces and how it is thought of in users' heads. These findings have motivated the design of an application that tries to mimic the latter and adapts to current computer activities.
In this paper we present an approach to contextual search, based on the automatically extracted metadata from visited documents. User model represents user's interests as a combination of tags, keywords and named entities. Such user model is further enhanced by automatically detected communities of similar users, based on the similarities of their models. The user may belong to multiple communities, each representing one of her possibly many personas -- roles or stereotypes, facets of her interests. We discuss further possibilities of using this model to bring more fine-grained contextualization and search improvement by using short contexts.
This work exploits folksonomy for building User Interest Profile (UIP) based on user's search history. UIP is an indispensable source of knowledge which can be exploited by intelligent systems for query recommendation, personalized search, and web search result ranking etc. A UIP consist of a clustered list of concepts and their weights. We show how to design, implement, and visualize such a system, in practice, which aids in finding interesting relationships between concepts and detect outliers, if any. The experiment reveals that UIP not only captures user interests but also its context and results are very promising.
Adaptation on public displays brings certain advantages and risks. Due to the implicit nature of adaptation, the users often miss the causality behind the adaptive behavior. Moreover, a high degree of autonomy in adaptive displays may leave the users with the feeling of control loss. Limited amount of transparency and controllability leads to the loss of user trust. As a result, the users feel insecure, frustrated, and are likely to abandon the system. The research goal of this work is to optimize the system actions in a ubiquitous display environment, in order make adaptation design transparent, controllable, and thus trustworthy. By means of a decision-theoretic approach the user trust can be assessed in different trust-critical contexts. The contexts describe the changes in the environment that call for adaptation: privacy of content, social setting, and accuracy of knowledge. The generated decisions enable the system to maintain trust and keep interaction comfortable.
The goal of this research is to investigate the effects of empathy and adaptive behaviour in long-term interaction between social robots and users. To address this issue, we propose an action selection mechanism that will allow a social robot to chose adaptive empathic responses, in the attempt to keep users engaged over several interactions.
The rise of socio-computational systems such as collaborative tagging systems, which rely heavily on user-generated content and social interactions, changed our way to learn and work. This work aims to explore the potentials of those systems for supporting knowledge work in organizational and scientific domains. Therefore, a user modeling approach will be developed which enables personalized services to shape the content towards individual information needs of novice, advanced and experienced knowledge workers. The novelty of this approach is a modeling strategy which combines user modeling characteristics from distinct research areas, the emergent properties of the socio-computational environment as well as non-invasive knowledge diagnosis methods based on the user's past interaction with the system.
Individuals with learning disabilities are excluded from the information and knowledge society because information present in such media, as well as software that enables access to it, does not meet their communication and accessibility requirements. To improve this situation, we have developed and evaluated an Instant Messaging (IM) service and client based on a pictographic communication system, and with a user interface designed taking into account their accessibility requirements. But the evaluation with individuals with learning disabilities has pointed out the need to take into account the great communications and computer skills diversity, even in groups with similar disability levels. Therefore, in this paper we present our plans to model the communication and accessibility requirements of individuals with learning disabilities in order to develop a mechanism to automatically personalize the IM client user interface and adapt it to their needs.
The dissertation project FamCHAI aims at creating a 'calendar companion' system in the form of a bidirectionally natural-language interactive scene with a virtual agent, and exploring the effects of adaptation of the agent to specific users both in terms of the support given (i.e. giving options the user likes) and in communication (i.e. presentation in a form the user prefers, and learning their idiosyncrasies for better understanding). Harnessing these models, interactions will grow steadily more effective, comfortable and natural for users.
Endowing systems with abilities to assess a user's mental state in an operational environment could be useful to improve communication and interaction methods. In this work we seek to model user mental workload using spectral features extracted from electroencephalography (EEG) data. In particular, data were gathered from 17 participants who performed different cognitive tasks. We also explore the application of our model in a non laboratory context by analyzing the behavior of our model in an educational context. Our findings have implications for intelligent tutoring systems seeking to continuously assess and adapt to a learner's state.
The open nature of exploratory learning leads to situations when feedback is needed to address several conceptual difficulties. Not all, however, can be addressed at the same time, as this would lead to cognitive overload and confuse the learner rather than help him/her. To this end, we propose a personalised context-dependent feedback prioritisation mechanism based on Analytic Hierarchy Process (AHP) and Neural Networks (NN). AHP is used to define feedback prioritisation as a multi-criteria decision-making problem, while NN is used to model the relation between the criteria and the order in which the conceptual difficulties should be addressed. When used alone, AHP needs a large amount of data from experts to cover all possible combinations of the criteria, while the AHP-NN synergy leads to a general model that outputs results for any such combination. This work was developed and tested in an exploratory learning environment for mathematical generalisation called eXpresser.
Assessing a learner's mastery of a set of skills is a fundamental issue in intelligent learning environments. We compare the predictive performance of two approaches for training a learner model with domain data. One is based on the principle of building the model solely from observable data items, such as exercises or test items. Skills modelling is not part of the training phase, but instead dealt with at later stage. The other approach incorporates a single latent skill in the model. We compare the capacity of both approaches to accurately predict item outcome (binary success or failure) from a subset of item outcomes. Three types of item-to-item models based on standard Bayesian modeling algorithms are tested: (1) Naive Bayes, (2) Tree-Augmented Naive Bayes (TAN), and (3) a K2 Bayesian Classifier. Their performance is compared to the widely used IRT-2PL approach which incorporates a single latent skill. The results show that the item-to-item approaches perform as well, or better than the IRT-2PL approach over 4 widely different data sets, but the differences vary considerably among the data sets. We discuss the implications of these results and the issues relating to the practical use of item-to-item models.
Virtual world user models have similarities with hypertext system user models. User knowledge and preferences may be derived from the locations users visit or recommend. The models can represent topics of interest for the user based on the subject or content of visited locations, and corresponding location models can enable matching between users and locations. However, virtual worlds also present challenges and opportunities that differ from hypertext worlds. Content collection for a cross-world search and recommendation service may be more difficult in virtual worlds, and there is less text available for analysis. In some cases, though, extra information is available to add to user and content profiles enhance the matching ability of the system. In this paper, we present a content collection system for Second Life and OpenSimulator virtual worlds, as well as user and location models derived from the collected content. The models incorporate text, social proximity, and metadata attributes to create hybrid user models for representing user interests and preferences. The models are evaluated based on their ability to match content popularity and observed user behavior.
Food and diet are complex domains for recommender technology, but the need for systems that assist users in embarking on and engaging with healthy living programs has never been more real. One key to sustaining long term engagement with eHealth services is the provision of tools, which assist and train users in planning correctly around the areas of diet and exercise. These tools require an understanding of user reasoning as well as user needs and are ideal application areas for recommender and personalization technologies. Here, we report on a large scale analysis of real user ratings on a set of recipes in order to judge the applicability and practicality of a number of personalization algorithms. Further to this, we report on apparent user reasoning patterns uncovered in rating data supplied for recipes and suggest ways to exploit this reasoning understanding in the recommendation process.