Rec-clusterGCN: An Efficient Graph Convolution Network for Recommendation

With more Graph Convolutional Network (GCN) models applied to recommendation tasks, graph data of tremendous scalability becomes much harder to train for existing models. Many existing works of adapting GCN models to recommendation tasks often try to improve accuracy by aggregating messages from all high-order neighboring nodes, which can result in insufficient computational efficiency. To tackle with data with large scalability, we propose a novel recommendation algorithm-Rec-clusterGCN, which utilizes the cluster structure of graphs, and is suitable for SGD-based training. It goes as follows: at every step, a dense subgraph of user-item interaction will be constructed, and each node will only aggregate messages from neighboring nodes within the subgraph. Next, our model accepts the LightGCN model structure. It can reduce a huge amount of computational time cost and make training of larger graph possible without sacrificing too much accuracy. In addition, by using our node enhancement technique, the performance of Rec-clusterGCN is further improved. The experimental results also indicate that the proposed algorithm outperforms most baseline algorithms. Significant improvement has been made in computational efficiency (on average about 30.0% relative improvement for different layers in the Gowalla dataset).