Double negative sampled graph adversarial representation learning with motif-based structural attention network

In recent years, graph neural networks have achieved remarkable performance in various downstream tasks by aggregating node neighborhoods hierarchically. However, prior methods usually treat neighboring nodes equally based solely on graph structure or rely on feature similarity for neighborhood aggregation, thereby underutilizing the integrated structural and attribute information. Inspired by the adversarial training mechanism, we propose a Double Negative sampled generative Adversarial network with Motif-based Structural Attention Network ( DNA-MSAN ), a novel framework for graph representation learning. Our method enhances the discriminator's capability to distinguish between positive and negative samples by preserving the original graph's motif-based higher-order structure and effectively learning node attribute information. Specifically, we first introduce an attention network based on the higher-order connectivity of motifs, which preserves the higher-order connectivity by perceiving motifs in the graph and using a normalized motif adjacency matrix for neighborhood aggregation. Next, we design a double negative sampling strategy, which cleverly introduces a negative sampling approach based on the current parameters of the discriminator to obtain additional negative samples, serving as useful supplementary samples to those generated by the generator. For the generator, we implement a local graph softmax that significantly optimizes the time complexity of traditional graph softmax by restricting the depth of breadth-first search traversal, leading to marked improvements in model performance. We evaluate DNA-MSAN on link prediction, node classification, and visualization tasks across five datasets, demonstrating substantial advancements in performance across multiple downstream tasks.