S-Walk: Accurate and Scalable Session-based Recommendation with Random Walks

Session-based recommendation (SR) aims at predicting the next items from a sequence of the previous items consumed by an anonymous user. Most existing SR models focus only on modeling intra-session characteristics, but neglect to consider inter-session relationships of items, helpful for improving the accuracy. Another critical aspect of recommender systems is computational efficiency and scalability, considering practical concerns in commercial applications. In this paper, we propose the novel Session-based Recommendation with Random Walk, namely S-Walk. Specifically, S-Walk can effectively capture both intra- and inter-session correlations on items by handling high-order relationships across items using random walks with restart (RWR). At the same time, S-Walk is highly efficient and scalable by adopting linear models with closed-form solutions for transition and teleportation matrices to formulate RWR. Despite its simplicity, our extensive experiments demonstrate that S-Walk achieves comparable or state-of-the-art performances in various metrics on four benchmark datasets. Moreover, the learned model by S-walk can be highly compressed without sacrificing accuracy, achieving two or more orders of magnitude faster inference than existing DNN-based models, particularly suitable for large-scale commercial systems.

• Data Mining and Modeling

• Machine Intelligence