Temporal Item Embedding with Static Similarity Regularization for Sequential Recommendation

Recommender systems have attracted a significant amount of research interests in recent years. Traditional methods such as content-based approaches and collaborative filtering approaches mainly focus on modeling the general user preference by using the user's whole purchase history. However, except for user preference, the sequential information should also be considered because human behavior exhibits sequential patterns. The methods considering both sequential item relationship and user preference are called sequential recommendation methods, which are mostly base on Markov Chains. Due to the difficulty of parameter estimation, most prior work only considers the latest interaction, which is insufficient according to our observations. To that end, in this paper, we propose a temporal item embedding method based on word2vec framework to model long purchase history, with each purchased item regarded as a word in sentence. Meanwhile, inspired by the collaborative filtering in traditional recommendations, we assume that similar items should have similar embeddings and propose static similarity regularization (SSR) in recommendation. The regularized item embedding can capture user preference and sequential item relationship simultaneously. Experiments on real-world datasets show that the proposed approach outperforms a spectrum of state-of-the-art algorithms.