Potential-Intent Graph Collaborative Filtering

Graph Convolutional Networks (GCNs) have exhibited significant efficacy in the representation of users and items within recommender systems. This exceptional performance is largely attributed to their ability to aggregate collaborative information from higher-order neighbors, thereby improving the optimization of embedding representations. However, with GCN-based models, node embeddings become similar when stacking too many layers, making them indistinguishable, which is known as the over-smoothing problem. We believe that an important element of the over-smoothing problem is that the multiple potential intentions a user may possess—such as merely passing time or selecting a gift for another individual—during the convolution of the multi-layer graph. For this reason, we propose a Potential-Intent Aware GCN-based recommendation model. This model effectively captures multiple potential intentions of users and constructs a set of sub-graphs comprising users with analogous interests. Then, the model performs higher-order graph convolution within these subgraphs. To evaluate the performance of the model, we conduct extensive experiments on three benchmark datasets, and the results indicating that our model outperforms existing GCN-based recommendation models.