基于随机游走的图扩散模型

A Graph Diffusion Model Based on Random Walk

图注意力网络(Graph Attention Networks,GAT)通过汇聚相邻节点计算中心节点特征时,缺少图的结构信息且没有利用高阶邻域节点。针对此问题,提出一种采用随机游走策略的图扩散模型。该模型通过随机游走访问邻域内的节点以提取结构信息,并通过设置游走时的重启参数和迭代次数,控制邻域范围以提取局部和全局节点信息,再通过图注意力的加权求和机制对特征进行转换。该模型在3个引文图数据集上进行实验比较,比传统GAT模型的准确率平均提升了1.1%,证明了随机游走策略在捕获节点结构信息方面发挥了重要作用。

Graph Attention Networks(GAT)lack structural information and do not utilize high-order neighboring nodes when aggregating features of central nodes by computing adjacent node contributions.To address this issue,this paper proposes a graph diffusion model employing a random walk strategy.This model accesses neighboring nodes through random walk to extract structural information.By controlling the neighborhood scope to extract local and global node information through the setting of restart parameters and iteration counts during walks,it utilizes a weighted sum mechanism of graph attention to transform features.Experimental comparisons on three citation graph datasets demonstrate an average accuracy improvement of 1.1%over the traditional GAT model,demonstrating the important role of the random walk strategy in capturing node structural information.