MPGAAN: Effective and Efficient Heterogeneous Information Network Classification

Abstract

In this paper, we propose a novel Graph Neural Network (GNN) model named "Meta-Path Guided Attention Aggregation Network" (MPAAGN), which is specifically designed for graph neural network classification algorithms based on attribute information aggregation. MPAAGN combines the advantages of Meta-Paths, GraphSAGE, and GAT (Graph Attention Networks) to deal with the node classification problem in heterogeneous information networks. The core idea of MPAAGN is to use meta-paths to define higher-order relationships between nodes in heterogeneous information networks to guide neighbor selection, and dynamically assign weights to different neighbors through the attention mechanism of GAT, so as to reflect their relative importance when aggregating neighbor information. At the same time, the model borrowed the neighbor sampling strategy of GraphSAGE to deal with the computational efficiency problem of large-scale graph data. The innovation of MPAAGN model is that it can effectively integrate the structure and attribute information of heterogeneous information networks, weight aggregate neighbor information through the attention mechanism, and capture high-order association by using meta-path, which is suitable for node classification tasks on large-scale graph data. The superior performance of the model in dealing with heterogeneous information network classification problems is proved by experiments, which provide a new research direction and practical tool for the field of graph neural networks.