Feature Mining and Node Prediction for Complex Correlation Datasets Based on Graph Neural Networks
Complex correlation datasets widely exist in social networks, citation networks, and biological systems, where traditional machine learning methods fail to effectively capture implicit topological correlation and high-dimensional feature interactions. Graph Neural Networks (GNNs) have unique advantages in processing non-Euclidean graph-structured data, which can realize adaptive feature mining and end-to-end node prediction. This paper proposes a multi-layer graph convolution feature mining model with adaptive neighbor aggregation, aiming at the problems of incomplete feature extraction and weak topological dependency perception in traditional GNNs for complex correlation datasets. Firstly, the graph structure normalization and feature dimensionality reduction are performed on the input complex graph data. Secondly, an improved graph convolution aggregation formula is constructed to fuse node attribute features and topological structure features hierarchically, and a residual connection mechanism is introduced to solve the gradient disappearance problem of deep graph networks. Finally, node classification prediction tasks are realized based on mined deep graph features. Comparative experiments are conducted on six classic complex graph benchmark datasets including Cora, Citeseer, Pubmed, Cornell, Texas, and Wisconsin. The experimental results show that the proposed model achieves average prediction accuracy improvements of 3.2%–8.7% compared with baseline models such as GCN, GAT, and GraphSAGE, which verifies the effectiveness and superiority of the proposed method in complex correlation data feature mining and node prediction.
