RCENR: A Reinforced and Contrastive Heterogeneous Network Reasoning Model for Explainable News Recommendation

Existing news recommendation methods suffer from sparse and weak interaction data, leading to reduced effectiveness and explainability. Knowledge reasoning, which explores inferential trajectories in the knowledge graph, can alleviate data sparsity and provide explicitly recommended explanations. However, brute-force pre-processing approaches used in conventional methods are not suitable for fast-changing news recommendation. Therefore, we propose an explainable news recommendation model: the Reinforced and Contrastive Heterogeneous Network Reasoning Model for Explainable News Recommendation (RCENR), consisting of NHN-R-2 and MR&CO frameworks. The NHN-R-2 framework generates user/news subgraphs to enhance recommendation and extend the dimensions and diversity of reasoning. The MR&CO framework incorporates contrastive learning with a reinforcement-based strategy for self-supervised and efficient model training. Experiments on the MIND dataset show that RCENR is able to improve recommendation accuracy and provide diverse and credible explanations.