Mixed Similarity Diffusion for Recommendation on Bipartite Networks

In recommender systems, collaborative filtering technology is an important method to evaluate user preference through exploiting user feedback data, and has been widely used in industrial areas. Diffusion-based recommendation algorithms inspired by diffusion phenomenon in physical dynamics are a crucial branch of collaborative filtering technology, which use a bipartite network to represent collection behaviors between users and items. However, diffusion-based recommendation algorithms calculate the similarity between users and make recommendations by only considering implicit feedback but neglecting the benefits from explicit feedback data, which would be a significant feature in recommender systems. This paper proposes a mixed similarity diffusion model to integrate both explicit feedback and implicit feedback. First, cosine similarity between users is calculated by explicit feedback, and we integrate it with resource-allocation index calculated by implicit feedback. We further improve the performance of the mixed similarity diffusion model by considering the degrees of users and items at the same time in diffusion processes. Some sophisticated experiments are designed to evaluate our proposed method on three real-world data sets. Experimental results indicate that recommendations given by the mixed similarity diffusion perform better on both the accuracy and the diversity than that of most state-of-the-art algorithms.