基于图多层感知机的节点分类算法

多数图神经网络(Graph Neural Networks, GNN)通过设计复杂的节点信息传递和聚合方式，以提升节点分类等图分析任务的实验表现，而本文提出了一种无需信息传递和聚合的图多层感知机(Multi-Layer Perceptron, MLP)模型A&T-MLP,利用属性和拓扑信息引导的对比损失来增强模型表征能力。A&T-MLP首先使用属性矩阵和邻接矩阵计算节点间的属性和拓扑相似度；然后使用基于相似度信息引导的对比损失，增大特征空间中相似节点的一致性和不相似节点的差异性；最后构建多层感知机模型并引入交叉熵损失进行端到端训练。在节点分类任务中，A&T-MLP表现优于基线模型，Wikipedia数据集上的Micro-F1和Macro-F1相较GNN模型图卷积网络(Graph Convolutional Network, GCN)提升了15.86%和13.64%。实验结果表明，A&T-MLP能够通过对比损失保留丰富原始图的信息，提升模型性能。此外，A&T-MLP在处理拓扑信息不准确的图数据时具有较为明显的优势，即使在缺失80%拓扑信息的极端情况...

Node Classification Based on Graph Multi-Layer Perceptron

Many Graph Neural Networks (GNN) improve the performance of graph analysis tasks,such as node classification,by designing efficient information propagation and aggregation methods.However,this paper proposes a graph Multi-Layer Perceptron (MLP) model A&T-MLP does not rely on information propagation and aggregation.It enhances its representation ability using through the utilization of contrastive loss guided by attribute and topology information.Firstly,attribute matrix and adjacency matrix were are used to calculate the attribute and topology similarity between nodes.Secondly,the contrastive loss guided by similarity information is applied to increase the consistency of similar nodes and the difference of dissimilar nodes in feature space.Thirdly,the multi-layer perceptron is constructed,and the end-to-end training process incorporates the cross-entropy loss function.A&T-MLP outperforms the baselines in the node classification.Micro-F1 and Macro-F1 increased by 15.86% and 13.64% respectively,compared to the GNN model Graph Convolutional Network (GCN) on Wikipedia dataset.The results show that A&T-MLP effectively preserves richer graph information,leading to improved performance.Besides,the proposed method exhibits notable advantages in inaccurate topology.Even in the extreme case where 80% of edges are missing ,A&T-MLP can still superior against baselines.