Residual Graph Convolution Collaborative Filtering with Asymmetric neighborhood aggregation

Due to the superior performance of graph convolutional networks (GCNs) in feature extraction and representation, researchers have introduced GCNs to collaborative filtering (CF) to improve the accuracy of recommendations. However, existing GCN-based CF models still have three shortcomings: degrading the model performance due to symmetric normalization in neighborhood aggregation, underutilizing the node (user or item) representations at lower layers, and failing to alleviate the over-smoothing effect in graph convolution operations. We present a Residual Graph Convolution Collaborative Filtering model with Asymmetric neighborhood aggregation (ARGCCF) to tackle the above shortcomings. Firstly, ARGCCF utilizes asymmetric normalization in neighborhood aggregation, enabling better preservation of high-degree node information for more precisely learning node representations. Secondly, the propagation of ARGCCF embeddings proceeds in two steps. During the initial propagation, a residual connection is established between the node representation of the current layer and the previous layer to utilize the low-level layer’s node representation fully. The embeddings generated after the initial propagation are propagated again in the same graph convolutional layer, which obtains more refined embedding representations to alleviate the over-smoothing problem. Finally, experimentations on five truthful datasets exhibit that ARGCCF has better-improved performance over several mainstream methods.