Improving the accuracy and diversity of personalized recommendation through a two-stage neighborhood selection

Abstract

Collaborative Filtering remains the most widely used recommendation algorithm due to its simplicity and effectiveness. However, most studies addressing the trade-off between accuracy and diversity in collaborative filtering recommendation algorithms focus solely on optimizing the recommendation list, often neglecting users’ diverse demands for recommendation results. We propose a new user-based Two-Stage collaborative filtering method for Neighborhood Selection (TSNS) that considers both the similarity between users and the dissimilarity between neighbors in the neighborhood selection phase. Firstly, we define the user’s preference value for the attributes of evaluated items and determine the range and ranking of user preferences. Then, we construct a preference heterogeneity model to evaluate preference differences among users and obtain a preference heterogeneity matrix based on the range and ranking of preferences. Finally, to effectively ensure recommendation accuracy and diversity, we adopt a two-stage neighborhood selection method to identify a group of neighbors that are internally dissimilar but similar to target users. Deep representation learning methods can also be incorporated into this framework to calculate user similarity in the first stage. Experimental results on two datasets show that our proposed method outperforms the benchmark method, including those using deep learning, in terms of comprehensive performance. Our approach offers new insights into improving the accuracy and diversity of personalized recommendations.