Abstract: Aiming at  the problems that graph attention networks were weak in handling noisy and temporal data, as well as gradient explosion and oversmoothing after stacking multiple layers, we proposed an improved graph attention network model. Firstly, we used  the Squeeze-and-Excitation module to pay different levels of attention to the feature information of the sample input data to enhance the model’s ability to handle noise. Secondly, the temporal sequence of the data was extracted by using the multi-head attention mechanism, which weighted and summed each sequence in the sequence data relative to the other sequences. Thirdly,  the node features extracted from the graph attention network were spliced with the degree centrality of the nodes to obtain the local features of the nodes, and the global features of the nodes were extracted by using global average pooling. Finally, the two were fused to obtain the final feature representation of the nodes, which enhanced the representational ability of the model. In order to verify the effectiveness of the improved graph attention network, the improved graph attention network model was compared with LSTM, GRU and GGNN models. The experimental results show that the prediction effect of the model has been effectively improved, with higher prediction accuracy.

Key words: graph attention network, multi-head attention, node degree centrality, global average pooling