Adaptive Neighbors Graph Convolutional Networks with Transfer Entropy Optimization

Graph Convolutional Networks (GCN) perform well in various graph neural network methods, particularly on homogeneous graphs. However, their performance on heterogeneous graphs has been less consistent. Many works attempt to modify the GCN architecture to address the challenges of heterogeneous graphs. However, these approaches often pay no attention to the nature of node interaction, specifically the exchange of information. In this paper, we propose an Adaptive Neighbors Graph Convolutional Networks with Transfer Entropy Optimization (ANGCN) to handle heterogeneous graphs. Transfer entropy quantifies directed information flow between two nodes over time. We first use transfer entropy to adjust the weight of information from neighboring nodes dynamically, and second, we apply it as a regularization term to constrain semantic information propagation. This process can reduce the influence of neighbors with more robust information transmission while enhancing the contribution of those with weaker transmission. Experimental results demonstrate that ANGCN outperforms state-of-the-art (SOTA) baseline models.