A Generalized Aggregation Method for Message Passing Graph Neural Networks

Graph Neural Networks (GNNs) have emerged as one of the most powerful tools for graph structure modeling. They have been successfully applied to solve different tasks in various domains and have gained impressive performances. Most GNNs are based on Message-Passing Neural Networks (MPNN), in which the representation update of a node is done iteratively. In each iteration, the update of a node representation only involves the information from its neighbors, leaving distant nodes untouched. Hence, it might not capture sufficient information. Here, we assume that the combination of information from neighboring and distant nodes whose high similarities with the current node measured by graph node kernels can improve the performance of the MPNN-based models. Therefore, we propose a generalized aggregation method that improves the performances of existing MPNN-based models. The evaluation results from various settings using different datasets and MPNN-based models confirm the potential of our proposed method.