Micro-behaviour with Reinforcement Knowledge-aware Reasoning for Explainable Recommendation

Existing practical recommendation scenarios involve multiple micro-behaviour user-item interactions, such as clicks, page views, add-to-favourites, and purchases, which provide fine-grained and a better in-depth understanding of the user's preference. Furthermore, some recommendation methods have incorporated item knowledge into the micro-behaviour of user-item interaction. Although some have proved effective, two insights are often neglected. First, they fail to combine micro-behaviour with the relation of the knowledge graph (KG), and the semantic relationship between micro-behaviour and relation is not captured. Second, they do not provide explicit reasoning for micro-behaviour from user-item interaction data. These insights motivated us to propose a novel model of Micro-behaviour with Reinforcement Knowledge-aware Reasoning for Explainable Recommendation (MBKR), which incorporates micro-behaviour and the KG into reinforcement learning for explainable recommendation. Specifically, the model learns the behaviour by user-item propagation and the relation from the KG and combines the two to calculate the behavioural strength to mine user's interests. In addition, we designed a Shawo-relational path that combines recommendation and interpretability by providing rational paths; these paths capture the semantics of behaviours and relations. Finally, we extensively evaluated our method on several large-scale benchmark datasets, and the results indicate that the proposed method is more effective in providing recommendations than state-of-the-art methods.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Existing recommendation methods neglect the relationship between micro-behavior and knowledge graph and the explicit reasoning for user-item interaction data. This paper proposes a model that incorporates micro-behavior and knowledge graph into reinforcement learning for explainable recommendation, achieving better recommendation results.