WARM: An Interpretability Module with Weighted Association Rule Mining for Recommendation Systems

Recommendation Systems interpretability has attracted more attention because it can increase users’ confidence in the Recommendation Systems. Most of the existing explainable recommendation models consider the interpretability of the recommendation methods, which is coupled to the recommendation model, lack of flexibility and generalization. To address the above issue, we propose an interpretability module for Recommendation Systems based on weighted association rule mining (WARM). The module is decoupled from the recommendation model and could be applied to the different latent factor models. WARM is improved on Apriori algorithm. Compared to Apriori algorithm, WARM converts rating into weight and make use of the weight in the association rules mining process, rather than just frequency, which can find more effective association rules to provide explanations for recommendations. Moreover, we are concerned with association rules of size greater than 2, which transforms the explanation style from the item-level to the theme-level, enhancing the persuasivity of using association rules to explain recommendations. We also describe the entire process of the interpretability module, explaining recommended model predictions with weighted association rules generated by WARM. Extensive experiments over real-world datasets demonstrate the effectiveness of WARM. The source code of WARM is available at https://github.com/zhuhaif/WARM .