Abstract:

In addressing the issue of relatively low accuracy in node classification tasks on heterophily graphs using methods such as MLP and GCN， a Graph Neural Network based on Similarity Random Walk Aggregation （SRW-GNN） was proposed. To address the impact of heterophily on node embeddings， SRW-GNN employs the similarity between nodes as probabilities for conducting random walks. The sampled paths serve as the neighborhood， enabling the model to gather more homophily-based information. To address the issue of insensitivity to node order in most existing graph neural network （GNN） aggregators， a path aggregator based on recurrent neural networks （RNNs） was introduced to simultaneously extract features and order information of each node in the path. Furthermore， nodes exhibit varying preferences for different paths. To adaptively learn the importance of different paths in node encoding， an attention mechanism was employed to dynamically adjust the contributions of each path to the final embedding. Experimental results on several commonly used heterophily graph datasets demonstrate that the proposed SRW-GNN method achieves significantly higher accuracy compared to the methods such as MLP， GCN， H2GCN， HOG-GCN， validating its effectiveness in heterophily graph node classification tasks.

Key words: computer application technology, graph neural network, heterophily graph, node classification