DPGNN: Dual-Perception Graph Neural Network for Representation Learning

TL;DR

This work formalizes the limitations of traditional graph neural networks' single-space, iterative message passing and introduces a novel Dual-Perception Graph Neural Network (DPGNN) that employs node-to-step attention and dual-space aggregation. By learning soft-weighted adjacencies across multiple hop lengths and integrating topology and feature spaces, DPGNN captures both structural neighborhood information and feature-driven signals. Extensive experiments on six datasets demonstrate state-of-the-art performance and highlight the importance of node-specific message outputs, multi-space interaction, and robust training via random sparse graph augmentation. The proposed framework advances graph representation learning by enabling flexible, diversified message passing with practical robustness across varying graph topologies.

Abstract

Graph neural networks (GNNs) have drawn increasing attention in recent years and achieved remarkable performance in many graph-based tasks, especially in semi-supervised learning on graphs. However, most existing GNNs are based on the message-passing paradigm to iteratively aggregate neighborhood information in a single topology space. Despite their success, the expressive power of GNNs is limited by some drawbacks, such as inflexibility of message source expansion, negligence of node-level message output discrepancy, and restriction of single message space. To address these drawbacks, we present a novel message-passing paradigm, based on the properties of multi-step message source, node-specific message output, and multi-space message interaction. To verify its validity, we instantiate the new message-passing paradigm as a Dual-Perception Graph Neural Network (DPGNN), which applies a node-to-step attention mechanism to aggregate node-specific multi-step neighborhood information adaptively. Our proposed DPGNN can capture the structural neighborhood information and the feature-related information simultaneously for graph representation learning. Experimental results on six benchmark datasets with different topological structures demonstrate that our method outperforms the latest state-of-the-art models, which proves the superiority and versatility of our method. To our knowledge, we are the first to consider node-specific message passing in the GNNs.

Figures (14)

• Figure 1: The topological views of six graph datasets. For intuitive presentation, we adopt different colors to distinguish nodes of different classes and set edges colorless. The number after the graph dataset name indicates the number of connected subgraphs of the graph. For example, there are 438 connected subgraphs in Citeseer, while Flickr is a big connected graph. Graph datasets Citeseer, ACM and CoraFull have similar topological structures, which are mainly composed of a large connected subgraphs (the middle part) and some small connected sub-graphs or isolated nodes (the surrounding part), while the topological structures of graph datasets UAI2010, Flickr and BlogCatalog are denser.
• Figure 2: (a) The existing message-passing paradigm, akin to a star-shaped pattern, defines the breadth of each message-passing iteration. (b) Message transitivity, akin to a chain-shaped pattern, ensures the depth of message passing over multiple iterations. (c) The boosting breadth of incoming messages requires increasing the depth of message passing. (d) The message-passing paradigm is analyzed from three aspects: message retention, message input, and message output. (e) An example illustrating the heterogeneity of different message-passing spaces.
• Figure 3: The framework of DPGNN model. 1) KNN Module: which constructs a feature graph $\mathcal{G}^f$ by the similarity of node featureswang2020gcnliu2021self. 2) Soft-Weighted Adjacency Matrix Learning Module: which learns two soft-weighted adjacency matrices $\overline{\mathbf{M}}^t$ and $\overline{\mathbf{M}}^f$ based on the two message-passing space. 3) Dual-Perception Representation Learning Module: which captures the structural neighborhood information and the feature-related information simultaneously for each node. 4) Optimization Module: which defines the loss function for DPGNN.
• Figure 4: Soft-Weighted Adjacency Matrix Learning Module
• Figure 5: Dual-Perception Representation Learning Module