Expectation-Maximization Collaborative Filtering with Explicit and Implicit Feedback

Abstract

Collaborative Filtering (CF) is a popular strategy for recommender systems, which infers users’ preferences typically using either explicit feedback (e.g., ratings) or implicit feedback (e.g., clicks). Explicit feedback is more accurate, but the quantity is not sufficient; whereas implicit feedback has an abundant quantity, but can be fairly inaccurate. In this paper, we propose a novel method, Expectation-Maximization Collaborative Filtering (EMCF), based on matrix factorization. The contributions of this paper include: first, we combine explicit and implicit feedback together in EMCF to infer users’ preferences by learning latent factor vectors from matrix factorization; second, we observe four different cases of implicit feedback in terms of the distribution of latent factor vectors, and then propose different methods to estimate implicit feedback for different cases in EMCF; third, we develop an algorithm for EMCF to iteratively propagate the estimations of implicit feedback and update the latent factor vectors in order to fully utilize implicit feedback. We designed experiments to compare EMCF with other CF methods. The experimental results show that EMCF outperforms other methods by combining explicit and implicit feedback.