An Improvement of Graph Neural Network for Multi-behavior Recommendation

One of the most extensively adapted paradigms for building a recommender system is embedding users and items into a low-dimensional latent space based on their interactions. While traditional Collaborative Filtering is designed for only one type of user-item action, real-world scenarios observe multiple activities of a user such as browsing and favorites, which can serve as an effective enhancement to the method. In addition, early efforts towards multi-behavior recommendation have two main limitations: first, they fail to calculate the influence strength of users' behaviors on target behavior; second, they also ignore that behavior semantics. Therefore, taking advantage of Graph Neural Network, this work further improves the graph-based recommender system by adding residual blocks and a behavior-learnable weight for each user. The effectiveness of our model is confirmed by empirical results of a real-world e-commerce dataset. Our model outperforms the baseline in both Recall@k and NDGC@k statistics.