A Bayesian Heterogeneous Graph Neural Network for Relational Uncertainty

Heterogeneous graphs composed of different types of nodes and edges have a wide range of application scenarios in the real world, such as social networks, transportation networks and c-commerce networks, which has attracted the attention to researchers. Although existing heterogeneous graph neural networks achieved excellent performance in the tasks of node classification, node clustering and link prediction, they do not take the uncertainty of relationships in a heterogeneous graph into account. Moreover, the performance of these models is significantly reduced when perturbed by adversarial examples, which exposes the weakness of the model in robustness. The weak robustness limits the application of heterogeneous graph neural networks of the real world, because heterogeneous network is easily changed by the disturbance of uncertain factors. Therefore, it is of great practical significance to improve the robustness of the model. In this paper, we propose a Bayesian Heterogeneous Neural Network (BHNN) to reveal and solve the problem of the uncertainty of relationships in a heterogeneous graph and to improve the robustness of the model. Firstly, BHNN predefines different meta-paths from the domain knowledge of a heterogeneous graph. Then, it constructs meta-path neighbor graphs through meta-paths collections. Secondly, BHNN uses a random block model to model each meta-path neighbor graph which can be regarded as a realization from a random graph parametric family. Finally, it uses the Bayesian method to infer the joint posterior of the parameters and node labels of the random graph to predict the unknown labels. In this process, the meta-path neighbor graphsarc reconstructed to reduce the weak edges and spurious edges caused by the uncertain relationship, so as to get generated graphs that can better reflect the real connection relationship. These generated graphs with different structures from the meta-path neighbor graph can be regarded as adversarial graph data, whose structures are more relevant and contain more abundant information. Furthermore, BHNN samples the weight samples of the generated graphs to further enhance its robustness. Extensive experiments arc conducted on three benchmark data sets including ACM, DBLP and IMDB. Compared with the state-of-the-art models, the Micro-Fland Macro-Fl of BHNN is increased by an average of 1. 59% and 1. 36% respectively, which illustrate the effectiveness and advancement of BHNN. Meanwhile, in the node attack experiments, BHNN maintains the best performance with less degradation in comparison with other heterogeneous graph neural networks, which also verifies that BHNN achieves a better performance of robustness. © 2023 Science Press. All rights reserved.