Item Similarity Learning Methods for Collaborative Filtering Recommender Systems

As one of the most popular recommender technologies, Collaborative Filtering (CF) has been widely deployed in industry due to its simplicity and interpretability. However, it is facing great challenge to generate accurate similarities between users or items because of data sparsity. This will cause second-order error in the process of using weighted sum as prediction. To alleviate this problem, we propose several methods to learn more accurate item similarities by minimizing the squared prediction error. This optimization problem is solved using Stochastic Gradient Descent. A comprehensive set of experiments on two real-world datasets at error and classification metrics indicate that the proposed methods can achieve comparable or even better performance than other state-of-the-art recommendation methods of Matrix Factorization, and greatly outperform traditional item based CF method. Besides, the proposed methods inherit the interpretability of item based CF, which makes the recommended results more accessible compared to competing methods of Matrix Factorization.