FeRe: Exploiting influence of multi-dimensional features resided in news domain for recommendation

摘要

As a promising direction, personalized news recommendation plays an important role in helping readers find interesting news from a gigantic amount of news items. Diverse methods related to news recommendation have been proposed to provide readers with personalized suggestions. However, little research focuses on exploiting influence of domain-specific features resided in news domain for recommendation. The objective of this paper is to build a hybrid news recommendation model, which improves news recommendation performance and alleviates novelly defined sparsity problem through exploiting influence of multi-dimensional domain-specific features from news, and saves much time in finding interesting stories for readers. First of all, we use a domain-specific feature from news, i.e., trends (different news categories have different lifecycle and play different roles in acquiring user profiles), to model user preference. The proposed user preference model dynamically evolves over time, through which, a superior content-based algorithm is proposed. Then, we also employ a domain-specific feature called user sequence with temporal feature, where multi-dimensional factors including user sequence, behavior weight, and time are used to compute similarity between users. The proposed similarity computation strategy is independent of common behaviors such as co-clicking/co-reading. Therefore, based on the similarity, improved collaborative filtering algorithm alleviates sparsity problem to some extent in terms of novelly definted sparsity metric relevant to news data. Besides, we also incorporate general news features into content-based and collaborative filtering news recommendation. Thus, we make the best of multi-dimensional features resided in news for improving recommendation performance. To further improve effectiveness, we investigate the feasibility of combining above two methods into a hybrid model called Fere by employing a no-argument union strategy. We conduct experiments to evaluate the effectiveness and efficiency of Fere by utilizing real news dataset. Experimental results show that Fere outperforms individual methods and baseline hybrid approaches.