Fusing Directed Negative Samples into Graph Convolutional Networks for Enhancing Node Learning

Graph neural networks (GNNs) are powerful models for learning graph-structured data. In most graph neural networks, only the first-order neighbouring nodes, also known as positive samples, are utilized for message passing, disregarding numerous non-neighbouring nodes that can be considered negative samples. Existing negative sampling methods do not consider the directionality of negative sampling, which leads to the undesirable feature change of negative samples and affects graph learning. To address this limitation, we propose the Directed Negative Sample Fusion Network (DNSFN), which adds directed edges from negative samples to central nodes and performs one-way convolution. In this way, the central node only receives information from the negative samples, not send information to them. Experiments show that DNSFN improves node learning, reduces over-smoothing, and outperforms traditional methods and GCN on higher layers, providing a new perspective for graph neural network research.