Recommender systems with heterogeneous information network for cold start items

Recommender System has been widely adopted in real-world applications. Collaborative Filtering (CF) and matrix based approach has been the forefront for the past decade in both implicit and explicit recommendation tasks. One prominent challenge that most recommendation approach facing is dealing with different data quality conditions. I.e. cold start and data sparsity. Some model based CF use condensed latent space to overcome sparsity problem. However, when dealing with constant cold start problem, CF based approach can be ineffective and costly. In this paper, we propose MERec, a novel approach that adopts graph meta-path embedding to learn item/user features independently besides learning from user-item interactions. It allows unseen data to be incorporated as part of user/item learning process. Our experiments demonstrated a effective impact reduction in cold start scenario for both new and sparse dataset.