Adaptive multi-channel Bayesian Graph Neural Network

Recent years have seen a surge in interest in graph neural networks (GNNs) due to their superior performance in a range of graph and network mining applications. Graph embedding attempts to convert nodes in graph data to a low-dimensional vector representation by capturing the edges between them. However, because the bulk of GNNs currently use unstable graph structures, they perform well on graphs with a high degree of homogeneity and badly on those with a low degree of homogeneity. As a result, GNNs that rely solely on the original graph structure may give unsatisfactory results. In this research, we introduce an adaptive multi -channel Bayesian graph neural network (AMBGN1) for estimating new graph structures and adaptively fusing some depth-related information between the original topological structures. The key idea is to follow the GNN mechanism by estimating the ideal graph structure using Bayesian inference and extracting specific and common embeddings from estimated graph structures, topological structures, and their combinations. We are able to maximize both estimated graph structure learning and node embedding in an iterative framework by using the attention approach to learn the significant weights of these three node embeddings. Our extensive research on a number of benchmark datasets with varied degrees of homogeneity demonstrated that AMBGN reliably estimates the graph structure and effectively learns the most relevant node information in both graph representations, confirming the usefulness of AMBGN.