ALGCN: Accelerated Light Graph Convolution Network for Recommendation

Ronghai Xu, Haijun Zhao, Zhi-Yuan Li, Chang-Dong Wang

Abstract: Recently, Graph Convolutional Network (GCN) has been widely applied in the field of collaborative filtering (CF) with tremendous success, since its message-passing mechanism can efficiently aggregate neighborhood information between users and items. However, most of the existing GCN-based CF models suffer from low convergence rates during training, mainly because they follow the design of standard GCN using a simple uniform average to aggregate the neighborhood information. We also find that the scale of embedding across different layers oscillates. We argue that these issues can be alleviated by our proposed graph convolution framework, namely Accelerated Light Graph Convolutional Network (ALGCN). ALGCN mainly contains two components: influence-aware graph convolution operation and augmentation-free in-batch contrastive loss on the unit sphere. Empirical evaluations on three large and public datasets demonstrate that the proposed method achieves remarkable training speedups over LightGCN and substantially outperforms the state-of-the-art GCN-based CF models. Our method also shows a great improvement in long-tail recommendation.