Random Walk Graph Auto-Encoders With Ensemble Networks in Graph Embedding
Chengxin Xie, Xiumei Wen, Fanxing Meng, Hui Pang
Abstract
Recently graph auto-encoders have received increasingly widespread attention as one of the important models in the field of deep learning. Existing graph auto-encoder models only use graph convolutional neural networks (GCNs) as encoders to learn the embedding representation of nodes. However, GCNs are only suitable for transductive learning, have poor scalability and shallow models with a poor perceptual field, and have limitations in node feature extraction. To alleviate these problems, we propose to use an adaptive weight integration graph attention network (GAT) and GCN’s random walk graph auto-encoder (EGRWR-GAE) to better learn the embedding representation of nodes. There is a large amount of noise in the graph data, which interferes with feature extraction and the GAT model is sensitive to noisy data, we propose a random walk graph auto-encoder (EGSRWR-GAE) that integrates GAT, GCN, and self-supervised graph attention networks (SuperGAT) using adaptive weights. The effectiveness of our model is well demonstrated by three publicly available datasets (Cora, Citeseer, and Pubmed) with optimizations of up to 2.2% on the link prediction task and up to 12.9% on the node clustering task.