Meta-MSGAT: Meta Multi-scale Fused Graph Attention Network

In recent years, graph attention networks(GAT) have received extensive attention. According to the different attention extraction methods, GAT is mainly divided into two categories: hierarchical graph attention and structural information extraction of attention. However, they cannot achieve plug-and-play effects. Recent studies have found that the excellent performance of Transformer on different tasks lies not in the multi-head attention mechanism but in the structure itself. Inspired by the above, we propose a plug-and-play graph attention convolution structure: Meta Multi-scale Fused Graph Attention Network(Meta-MSGAT). Meta-MSGAT consists of multi-scale feature extraction, feature transformation and aggregation, and message-passing. The Multi-scale feature extraction is implemented using multiple channels. Feature transformation and aggregation consist of different convolution kernels to increase the diverse representation of features. The message-passing layer consists of a mixer that does not limit the specific network layer. To verify the effectiveness of our proposed structure, we performed extensive experiments on six datasets and achieved an improvement from 0.35% to 2.76% compared to baselines. In particular, we use MLP or GAT instead of the mixer for ablation experiments, and both achieve superior performance over baselines.