Structure-Aware Consensus Network on Graphs with Few Labeled Nodes

TL;DR

This work tackles semi-supervised graph node classification under very limited labeling, where traditional methods underutilize unlabeled data and graph structure. It introduces Structure-Aware Consensus Network (SACN), a single-branch model that performs strong-strong consensus learning between two strongly augmented views and augments this with a weak-to-strong pseudolabel supervision guided by a class-balanced strategy. The core innovations are a structure-aware consensus objective that combines $\ell_{cor}$ and $\ell_{de}$ through $\ell_{sacn} = \ell_{cor} + \lambda \ell_{de}$, and a class-aware pseudolabeling mechanism that supervises two strong views using high-confidence labels from a weak view via cross-entropy, i.e., $\ell_{w2s}$. Experiments on Cora, Citeseer, and PubMed show SACN achieving state-of-the-art performance at label rates as low as 0.5%–0.03%, with fewer parameters and improved efficiency due to neighborhood-aware correlations; results remain robust under imbalanced class distributions. These findings imply SACN’s practical impact for scalable, accurate semi-supervised node classification on large, skewed graphs.

Abstract

Graph node classification with few labeled nodes presents significant challenges due to limited supervision. Conventional methods often exploit the graph in a transductive learning manner. They fail to effectively utilize the abundant unlabeled data and the structural information inherent in graphs. To address these issues, we introduce a Structure-Aware Consensus Network (SACN) from three perspectives. Firstly, SACN leverages a novel structure-aware consensus learning strategy between two strongly augmented views. The proposed strategy can fully exploit the potentially useful information of the unlabeled nodes and the structural information of the entire graph. Secondly, SACN uniquely integrates the graph's structural information to achieve strong-to-strong consensus learning, improving the utilization of unlabeled data while maintaining multiview learning. Thirdly, unlike two-branch graph neural network-based methods, SACN is designed for multiview feature learning within a single-branch architecture. Furthermore, a class-aware pseudolabel selection strategy helps address class imbalance and achieve effective weak-to-strong supervision. Extensive experiments on three benchmark datasets demonstrate SACN's superior performance in node classification tasks, particularly at very low label rates, outperforming state-of-the-art methods while maintaining computational simplicity.The source code is available at https://github.com/kunzhan/SACN

Figures (3)

• Figure 1: Different self-supervised strategies. Attributed feature matrix $X$ and graph structure $A$ to the model, output prediction representation $Y$. (a) Previous simple pseudo supervision strategy. (b) View-consistency strategy. (c) the proposed SACN mainly uses a structure-aware consensus objective between strongly augmented views. Furthermore, we use a weak-to-strong supervision between $\hat{T}$ and $Y^{(v)}$. We use a class-aware strategy to generate pseudolabels $\hat{T}$, which is used to supervise the predictive representations ${Y}^{(1)}$ and ${Y}^{(2)}$.
• Figure 2: Structure-aware consensus network architecture. SACN is constructed on a two-layer graph attention network, consists of two augmented views and one normal view.
• Figure 3: $t$-SNE visualization of the learned representations from SACN algorithm on Coar, Citeseer and Pubmed separately during training, different categories of points are colored differently. In the above figures: (a) The input data X. (b) Feature after 100 epochs. (c) Feature after 200 epochs. (d) Feature after 400 epochs. With the increase of epochs, representations of the same class are close while those of different classes are far away gradually.