An Enhanced Graph Neural Network Based on the Tissue-Like P System

Abstract

Graph neural networks (GNNs) have achieved excellent performance in processing and analyzing graph-related tasks. Most of the message-passing GNN models adopt a neighborhood aggregation method, that is, update node embeddings by aggregating neighbor features. However, it is not always necessary to aggregate the information of all neighborhood nodes. The feature information of some neighbors is redundant or even harmful. In this paper, we propose an enhanced graph neural network based on the tissue-like P system (EGNN-P), which integrates the process of optimizing the aggregation of neighbor information into the tissue-like P system. The main idea is to reduce the aggregation of harmful information or redundant information by removing some edges connecting nodes in each training epoch. EGNN-P is a general framework that can be combined with many baseline GNN models on the task of node classification, such as GCN, GAT, and GraphSAGE. Experiments on three real-world citation network datasets demonstrate the novel method EGNN-P can effectively improve the performance the baseline GNN model in semi-supervised node classification.