Causal Behavior Pattern Inference for News Recommendation Through Multi-interest Matching

Xingming Chen, Wenqi Fan, Qing Li

Abstract: Personalized news recommendation is essential for helping users efficiently discover content aligned with their interests. Current methods primarily treat it as a sequential task, focusing on statistical correlations within news content, which hinders the inference of causal patterns in user behaviors. Unlike conventional sequential recommendation, users’ browsing behaviors are generally driven by diverse and evolving interests, leading to dissimilar adjacent news selections. Uncovering causal relationships among user interests can reveal the underlying behavioral patterns. For instance, users are more likely to browse lighter content after reading serious political news. Ignoring such causal patterns can result in recommendations that overly emphasize similar content. To address this, we propose Causal Behavior Pattern Inference (CBPI), a framework that models user behavior from a causal perspective. CBPI infers multiple latent interests and uncovers their causal structures, while dynamically adapting to changes in user preferences. By mapping interest-level causal relations to news-level interactions, CBPI offers a more accurate understanding of user preferences. Extensive experiments on real-world datasets show that CBPI outperforms state-of-the-art methods.

External IDs:dblp:conf/wise/ChenFL24