User Modeling and User-Adapted Interaction最新文献

Modern music streaming services rely on recommender systems to help users navigate within their large collections. Collaborative filtering (CF) methods, that leverage past user–item interactions, have been most successful, but have various limitations, like performing poorly among sparsely connected items. Conversely, content-based models circumvent the data-sparsity issue by recommending based on item content alone, but have seen limited success. Recently, graph-based machine learning approaches have shown, in other domains, to be able to address the aforementioned issues. Graph neural networks (GNN) in particular promise to learn from both the complex relationships within a user interaction graph, as well as content to generate hybrid recommendations. Here, we propose a music recommender system using a state-of-the-art GNN, PinSage, and evaluate it on a novel Spotify dataset against traditional CF, graph-based CF and content-based methods on a related song prediction task, venturing beyond accuracy in our evaluation. Our experiments show that (i) our approach is among the top performers and stands out as the most well rounded compared to baselines, (ii) graph-based CF methods outperform matrix-based CF approaches, suggesting that user interaction data may be better represented as a graph and (iii) in our evaluation, CF methods do not exhibit a performance drop in the long tail, where the hybrid approach does not offer an advantage.

Smart Product–Service Systems (S-PSS) represent an innovative business model that integrates intelligent products with advanced digital capabilities and corresponding e-services. The user experience (UX) within an S-PSS is heavily influenced by the customization of services and customer empowerment. However, conventional UX analysis primarily focuses on the design stage and may not adequately respond to the evolving user needs during the usage stage and how to exploit the data surrounding the use of S-PSS. To overcome these limitations, this article introduces a practical framework for developing Adaptive User Interfaces within S-PSS. This framework integrates ontologies and Context-aware recommendation systems, with user interactions serving as the primary data source, facilitating the development of adaptive user interfaces. One of the main contributions of this work lies on the integration of various components to achieve the creation of Adaptive User Interfaces for digital services. A case study of a smart device app is presented, to demonstrate the practical implementation of the framework, with a hands-on development approach, considering technological aspects and utilizing appropriate tools. The results of the evaluation of the recommendation engine show that using a context-aware approach improves the precision of recommendations. Furthermore, pragmatic aspects of UX, such as usefulness and system efficiency, are evaluated with participants with an overall positive impact on the use of the smart device.

Online news article recommendations are typically of the ‘more like this’ type, generated by similarity functions. Across three studies, we examined the representativeness of different similarity functions for news item retrieval, by comparing them to human judgments of similarity. In Study 1 ((N=401)), participants assessed the overall similarity of ten randomly paired news articles on politics and compared their judgments to different feature-specific similarity functions (e.g., based on body text or images). In Study 2, we checked for domain differences in a mixed-methods survey ((N=45)), surfacing evidence that the effectiveness of similarity functions differs across different news categories (‘Recent Events’, ‘Sport’). In Study 3 ((N=173)), we improved the design of Study 1, by controlling for how news articles were matched, differentiating between dissimilar news articles and articles that were matched on a shared topic, named entities, and/or date of publication, across ‘Recent Events’ and ‘Sport’ categories. Across all studies, we found that users mostly used text-based features (e.g., body text, title) for their similarity judgments, while BodyText:TF-IDF was found to be the most representative for their judgments. Moreover, the strength of similarity judgments by humans and similarity scores by feature-specific functions was strongly affected by how news article pairs were matched. We show that humans and similarity functions are better aligned when two news articles are more alike, such as in a news recommendation scenario.

In recent years, research on transformer-based recommender systems has attracted a lot of attention. Our work examines how BERT, a transformer-based contextual language model, can be applied to build a session-based news recommender system. The session-based approach aims to recommend news by creating profiles for users and items and recommending items accordingly to maximize session length. The proposed model, called SNRBERT (Session-Based News Recommender using BERT), is fine-tuned to estimate the relationship between each user and item in a given session based on the interactions between the user and the item during that session. We introduce this method to address the challenges of session-based news recommendation, particularly in maximizing session length and capturing user–item relationships effectively. Given the limited information available about user preferences in session-based scenarios, the model estimates a score based on the amount of time users spend on each item in each session. The news recommendation is then performed based on this score. On top of BERT, we employed an Bi-LSTM network in order to capture more accurate information regarding the order in which users interact with items during a given session. We compare our results with the state-of-the-art models by using commonly known metrics: MRR, HR, and NDCG on the Adressa dataset, one of the most comprehensive datasets publicly available. Our results show that our SNRBERT model achieves HR@10 of 0.688, MRR@10 of 0.315, and nDCG@10 of 0.338. These results demonstrate that SNRBERT outperforms state-of-the-art models in terms of MRR@10 and HR@10 metrics, showcasing its effectiveness in addressing the challenges of session-based news recommendation.

With the prominence of online social networks, news has become more accessible to a global audience. However, in the meantime, it has become increasingly difficult for individuals to differentiate between real and fake news. To reduce the spread of fake news, researchers have developed different classification models to identify fake news. In this paper, we propose a fake news detection system using a multilayer perceptron (MLP) model, which leverages linguistic and psychological features to determine the truthfulness of a news article. The model uses different features from the article’s text content to detect fake news. In the experiment, we utilize a public dataset from the FakeNewsNet repository consisting of real and fake news articles collected from PolitiFact and BuzzFeed. We perform a meta-analysis to compare our model’s performance with existing classification models using the same feature sets and evaluate the performance using the metrics such as prediction accuracy and F1 score. Overall, our classification model produces better results than existing baseline models, by achieving an accuracy and F1 score above 90 % and performs 3% better than the best performing baseline method. The inclusion of linguistic and psychological features with a deep neural network allows our model to consistently and accurately classify fake news with ever-changing forms of news events.

Physical inactivity is a public health issue. Mobile health interventions to promote physical activity often still experience dropout, resulting in people not adhering to the interventions. This paper aims to further improve mobile health apps with innovatively applied techniques from recommender system algorithms to increase personalization for physical activities and practical tips to reduce sedentary behavior. Personalization in our mobile health recommender is achieved with a seven-step algorithm: filtering on user profile (1), current weather and daylight (2), pre-filtering with a micro-profile on current mood and motivation (3), content-based recommendations using our own two datasets extended with 24 attributes (4), post-filtering on estimated current situation (5), adapting and gradually increasing duration and intensity (6), and generating just-in-time adaptive interventions (7). To analyze the effectiveness of steps 3, 4, and 5, a double-blind randomized controlled trial is conducted in which only the experimental group receives the three additional personalization steps, while the control group replaces these steps with a random selection. As such, the control group’s recommendations are still partly personalized with the other steps. Participants install the app on their Android smartphone and use the app for eight weeks, with a pretest and posttest questionnaire, and a follow-up after six months. The experimental group assigned significantly higher star ratings to the recommendations, and significantly higher momentary motivation for physical activities, tips, and manual user refreshes, compared to the control group. Additionally, there was less dropout and a significantly stronger increase in duration and intensity of the performed physical activities in the experimental group. Because the experimental group received the three additional personalization steps with micro-profiling, content-based recommender, and post-filtering on estimated situation, our results suggest that these three steps resulted in more personalized recommendations that motivate users more. Future research should aim to further improve personalization to increase the effectiveness of mobile health interventions and effectively motivate people to move more.

Recommender systems help people find relevant content in a personalized way. One main promise of such systems is that they are able to increase the visibility of items in the long tail, i.e., the lesser-known items in a catalogue. Existing research, however, suggests that in many situations today’s recommendation algorithms instead exhibit a popularity bias, meaning that they often focus on rather popular items in their recommendations. Such a bias may not only lead to the limited value of the recommendations for consumers and providers in the short run, but it may also cause undesired reinforcement effects over time. In this paper, we discuss the potential reasons for popularity bias and review existing approaches to detect, quantify and mitigate popularity bias in recommender systems. Our survey, therefore, includes both an overview of the computational metrics used in the literature as well as a review of the main technical approaches to reduce the bias. Furthermore, we critically discuss today’s literature, where we observe that the research is almost entirely based on computational experiments and on certain assumptions regarding the practical effects of including long-tail items in the recommendations.

Factual knowledge and procedural knowledge are knowing ‘That’ and ‘How,’ respectively, whereas conditional knowledge is the metacognitive knowledge of ‘When’ and ‘Why.’ As prior work has found that students with conditional knowledge spontaneously transferred such knowledge across intelligent tutoring systems, this work assesses the impact of metacognitive interventions on the knowledge transfer of factual and procedural students. Specifically, we used a between-subject, pre-/posttest design with factual and procedural students, each randomly assigned to either the example, nudge, practice, or control condition. The interventions taught how and when to use a backward-chaining (BC) strategy on a logic tutor that supports a default forward-chaining strategy. Meanwhile, conditional students received no interventions. Six weeks later, we trained all students on a probability tutor that only supports BC without interventions. Our results suggest that nudges for factual students and practice for their procedural peers are the key factors for catching up with conditional students on both tutors and for facilitating knowledge transfer from the logic to probability tutor. We discuss two potential complementary theories for our findings: a choice-based theory (from interventions to knowledge) and a metacognitive load-based theory (from knowledge to interventions). The choice-based theory maps the amount of choice in the interventions to knowledge types, while the metacognitive load-based theory associates knowledge types with the metacognitive load each intervention offers. Implications for practice are discussed.

Cold start and data sparsity are problems hindering the function of news recommender systems. Optimally serving first-time users through relevant news article recommendations is an application of these problems that have attracted scholars’ attention. Users’ goal commitment might be another solution that raise efficiency of information searching while it is understudied in previous research. Drawing from the results of 669 Amazon MTurk workers’ questionnaires, this experimental study explored solutions. We manipulated the relevance of news recommendations (high relevance vs. low relevance) and information behavior within a news portal, either scanning (via a list of news articles) or seeking (via a search query). We also measured an individual difference variable, goal commitment. Results indicated that higher relevance of recommendations and higher goal commitment lead to lower information overload, higher user satisfaction, and lower information anxiety. We also found interaction effects of goal commitment and content relevance on article selection, such that users will be likely to select more irrelevant articles in the low relevance condition rather than the high relevance condition even though they have a goal commitment and perceive higher information overload and information anxiety indirectly via selecting more irrelevant articles. Furthermore, people with high goal commitment were less anxious when they read fewer irrelevant articles in the news recommender systems. The study addressed the importance of considering the user-recommender interaction and the potential merits of considering users goal commitment in the news recommender system design. The research indicates integrating personal traits into state-of-the-art news recommender systems has the potential to significantly improve user experience. While this research suggests personal traits can mitigate the limitations of imperfect recommender systems, users can also curate or train these systems based on their goals to further enhance efficiency.

Learning can generally be categorised into three domains, which include cognitive (thinking), affective (emotions or feeling) and psychomotor (physical or kinesthetic). In the learner model, acknowledging the affective aspects of learning is important for a range of learner outcomes, including motivation, persistence, and engagement. Learners’ affective states can be detected using physical (e.g. cameras) and physiological sensors (e.g., EEG) in online learning. Although these detectors demonstrate high accuracy, they raise privacy concerns for learners and present challenges in deploying them on a large scale to larger groups of students or in classroom settings. Consequently, researchers have designed an alternative method that can recognise students’ affective states at any point during online learning from their interaction with a computer-based learning platform (i.e. intelligent tutoring systems) without using any sensors. Existing sensor-free affect detectors however, are less accurate and not directly generalisable to other domains and systems. This research focuses on developing generalisable sensor-free affect detectors to identify students’ frustration during online learning using machine learning classifiers. The detectors were built by identifying minimal optimal features associated with frustration from the high-dimensional feature space through a series of experiments on a real-world students’ affective dataset, which are generalisable across various learning platforms and domains. To evaluate their accuracy and generalisability, the detectors’ performance was validated on two independent datasets collected from different educational institutions. The experimental results show that cost-sensitive Bayesian classifiers can achieve higher affect detection accuracies with a small number of generalisable features compared to other classifiers.