Original link: http://vis.pku.edu.cn/blog/newsrecxplain/
The Internet has made it easier than ever to distribute news articles to more people, and people’s access to information is overflowing with data content and digital objects. Recommender systems (RSs for short) have become indispensable. However, RSs always lack transparency and diversity, and users have little control over “what is recommended”. In an era where reliance on digital information can have a huge impact on our lives, this has created a sense of mistrust in recommender systems. Although concepts such as explanations, control, and empowerment have been widely discussed in the field of RSs. However, little attention has been paid to how to use visualization to advance these concepts to improve the problems existing in existing recommender systems.
Researchers from Switzerland propose NewsRecXplain, which uses interactive and interpretable machine learning methods to support users in understanding, diagnosing, and improving personalized recommendation systems. NewsRecXplain’s interface allows users to interpret and customize news recommendations, and helps empower users to use recommender systems, enabling them to detect and improve filter bubble effects in information dissemination.
Users of NewsRecXplain are readers of online news, and their contributions mainly include:
(1) Understanding news recommendations through visual interpretation of recommended content;
(2) Guiding the recommendation model by configuring personalized personal profiles and controlling model parameters, that is, facilitating the customization of RSs;
(3) enable users to improve the filter bubbles in their reading of information by placing their customized models in the context of other potentially existing configured models;
Figure 1: The process of NewsRecXplain interacting with users
Its tasks mainly include:
(1) Explanations: Explore recommendation output features to improve users’ understanding of profile-driven decisions.
(2) Customization: select user profiles and control model parameters to control RSs.
(3) Empowerment: Associating the user’s customized model with other configurations, enabling them to control their own filter bubbles.
The paper uses the Microsoft News Dataset (MiND), which includes 45,463 news articles and 29,572 users. The authors of the dataset extracted any named entities (such as persons and organizations) mentioned in news articles and matched them to records in Wikidata. Wikidata provides information about relationships between entities, which is used to generate embeddings that encode connections between entities in a high-dimensional space.
The authors summarize RSs recommender systems into three main recommendation methods: collaborative filtering, content-based filtering, and popularity-based filtering, and use 2D visualization to project the internal structure of news article embeddings. NewsRecXplain also uses statistical visualizations to report the diversity of recommendations and the contribution of three different recommendation methods to the recommendation results.
Figure 2: Three main recommendation methods
The NewsRecXplain visualization workspace consists of three components: model configuration, explanation, and news recommendation. Method control noodles enable the user to control the impact of the recommended parameters. The shape of the memory function of the user’s viewing history may change as the user adjusts the parameters.
Figure 3, 4: NewsRecXplain system interface and custom panel
NewsRecXplain supports users to understand and explore their custom models and transform the models to their liking. Visualizations can support this process by providing visual metaphors for user preferences and other user models. NewsRecXplain is the first step towards using visualization in personalized news RS. Since this is a visualization-centric work, some underlying algorithmic solutions are still limited. . Likewise, the performance of the selected recommendation algorithms was not optimized or evaluated against the benchmarks of this work. In the future, it can improve the accuracy by the standard evaluation method of MiND challenge.
This article is transferred from: http://vis.pku.edu.cn/blog/newsrecxplain/
This site is only for collection, and the copyright belongs to the original author.