Interpretable User Behavioral Analysis and Personalized Recommendation with Side Information

Due to information overload, recommender system is developed and widely applied to better predict customer behaviors empowered by interaction data. In this paper, we develop an interpretable recommender system to improve click behavior prediction accuracy. Our proposed recommender system consists of i) designing a user-item interaction rating system defined as the weighted-sum of the multiple interaction variables, where the weight vector is regarded as hyperparameter, ii) using matrix factorization technique to capture the latent structure from the observable user-item interaction, and predict personalized ratings for items not shown to users, and iii) determining the weight vector by optimizing AUC which evaluates click behavior prediction accuracy on a validation set. To be specific, an algorithm with radial basis function surrogates is applied for hyperparameter optimization. We also develop a case study with user behavioral data from Netease music and observe a 15.5% improvement in AUC when incorporating multiple information sources.