Polarized Message-Passing in Graph Neural Networks

With the widespread application of graph-structured data in various fields, Graph Neural Networks (GNNs) have attracted significant attention as a powerful tool for analyzing graph data. However, existing GNNs primarily rely on neighborhood node similarity information when learning node representations, overlooking the potential of node differences. Recently, a novel “Polarized Message-Passing” (PMP) paradigm has emerged, injecting a fresh concept into GNN design.

Research Background: Traditional GNNs learn the representation of a target node by aggregating the features of its neighboring nodes, but they only consider the similarity between nodes and fail to fully utilize the rich information inherent in node differences. In reality, phenomena such as the “friendship paradox” and “influence imbalance” are prevalent in real-world graph data, reflecting the unique qualities of nodes in the graph. In light of this, this research proposes the PMP paradigm, aiming to simultaneously capture both node similarity and difference information, thereby enhancing the representation learning capability of GNNs.

Research Institutions: This research was jointly conducted by scholars from the Agency for Science, Technology and Research (A*STAR) in Singapore, Hong Kong Baptist University, Beijing University of Posts and Telecommunications, and Southwest University. The paper was published in the 2024 top artificial intelligence journal “Artificial Intelligence.”

Research Content: a) Research Process - For each layer of the GNN, PMP first constructs two learnable matrices quantifying the relevance and difference between nodes, respectively. - PMP then fuses the two matrices through an exponential operation to obtain a polarized message weight matrix, which is used to combine the features of neighboring nodes. - Finally, the representation of the target node is formed by aggregating its own features and the modulated features of its neighboring nodes using the weight matrix.

b) Main Results - Theoretical analysis proves that the PMP paradigm endows GNNs with stronger expressive power, effectively capturing node heterogeneity in graph data. - Three novel GNNs are proposed based on PMP: PMP Graph Convolutional Network (PMP-GCN), PMP Graph Attention Network (PMP-GAT), and PMP Graph PageRank Network (PMP-GPN). - On 12 real-world datasets and 5 downstream tasks, the proposed three PMP-GNNs outperform existing mainstream GNN models.

c) Research Significance - The PMP paradigm helps GNNs learn more expressive node representations by reasonably utilizing node difference information. - The novel PMP-GNN models demonstrate excellent performance in various application scenarios, opening new perspectives for graph data analysis. - This research provides a fresh approach to GNN design, promising to promote innovative applications of GNNs in complex graph data mining.

d) Research Innovation - For the first time, node difference information is integrated into the message-passing paradigm of GNNs. - The novel PMP paradigm is proposed to address the limitations of existing GNNs that solely utilize node similarity. - Three innovative GNNs based on PMP are invented, possessing stronger expressive power and robustness.

The PMP paradigm innovatively incorporates node difference information into the representation learning process of GNNs, expanding the application prospects of GNNs and promising to drive the theoretical and practical development of graph mining.