Heterogeneous Graph Neural Network based on Bandit Sampling for Attribute Completion

Graph neural networks (GNNs) have gained significant attention across diverse areas due to their superior performance in learning graph representation. Many models have emerged to handle heterogeneous graphs using GNNs and have achieved significant success. However, the challenge of incomplete attribute values remains prevalent, particularly in heterogeneous graphs where some types of nodes lack attributes. Existing approaches often separate the attribute completion from the graph learning process. This separation falls short of fully exploiting the rich information inherent in heterogeneous graphs. In this paper, we propose a methodology by using a heterogeneous graph neural network based on bandit sampling for attribute completion. Our approach consists of three key modules: topological embedding, adaptive node sampling, and representation learning. It first learns topological information, and then by integrating a modified multi-armed bandit algorithm, our proposed method adaptively samples informative nodes. This enhances the attribute completion process. Finally, it learns the final embedding of the heterogeneous graph using the graph with completed attributes. We conduct extensive experiments on three real-world datasets. Compared with the state-of-the-art heterogeneous GNN models on the node classification task and the node clustering task, our approach improves the performance by up to 2% and 2.2%, respectively. This shows that the completed attributes from various node types will be aggregated into the target node types, hence improving the model’s predictive performance.