Reinforced logical reasoning over KGs for interpretable recommendation system

Shirui Wang, Bohan Xie, Ling Ding, Jianting Chen, Yang Xiang

Abstract: In various domains, traditional recommendation systems have demonstrated significant benefits. However, their "black box" mechanisms have led to a crisis of trust among users. Interpretable recommendation systems have emerged as a solution by providing explanations for recommended items, thus enhancing transparency and user confidence. Another challenge to interpretable recommendation systems is data sparsity, which causes subpar recommendation performance. Addressing the challenges of model interpretability and data sparsity, this paper introduces the Knowledge Graphs-based Logic Reasoning Recommendation (KG-LRR) method, structured around an "encoder-decoder" architecture. The KG-LRR method tackles these issues by leveraging a knowledge graph for items to enhance the representation of users and items during the encoding process. It introduces a propositional logic reasoning model for decoding, rendering explanations in a more comprehensible manner. This dual approach ensures a balance between the recommendation system’s efficiency and interpretability. The KG-LRR method employs a neural network to simulate human-like propositional logical reasoning. This not only mitigates data sparsity issues but also explicates users’ interest in items. It provides deeper insights into users’ preferences and delivers robust interpretability. Experimental results across three public datasets-Yelp2018, Amazon-book, and Amazon-electronics-demonstrate that the KG-LRR model outperforms existing methods in terms of Recall and nDCG in top-k ranking recommendation scenarios. This validates its superior performance compared to prevailing interpretable recommendation techniques. In summary, the KG-LRR method offers a novel approach to enhance transparency and performance through an innovative "encoder-decoder" architecture. Its integration of knowledge graphs and propositional logic reasoning showcases promising outcomes in addressing current challenges within interpretable recommendation systems. Our code is available at https://github.com/siri-ya/KG-LRR.