Edge Contrastive Learning: An Augmentation-Free Graph Contrastive Learning Model

TL;DR

This work introduces Augmentation-Free Edge Contrastive Learning (AFECL), a graph representation learning framework that performs edge-edge contrast without data augmentations. By deriving edge embeddings directly from node embeddings produced by a multi-head GAT encoder, and by defining positives and negatives through local topology (edges adjacent to shared nodes vs. non-neighbor edges), AFECL leverages graph structure while maintaining computational efficiency. The approach achieves state-of-the-art results on link prediction and semi-supervised node classification with scarce labels across eight diverse datasets, and ablations confirm the efficacy of the edge-centric loss and sampling strategy. Overall, AFECL provides a scalable, topology-aware alternative to augmentation-heavy graph contrastive learning with strong practical impact on graph analysis tasks in low-label regimes.

Abstract

Graph contrastive learning (GCL) aims to learn representations from unlabeled graph data in a self-supervised manner and has developed rapidly in recent years. However, edgelevel contrasts are not well explored by most existing GCL methods. Most studies in GCL only regard edges as auxiliary information while updating node features. One of the primary obstacles of edge-based GCL is the heavy computation burden. To tackle this issue, we propose a model that can efficiently learn edge features for GCL, namely AugmentationFree Edge Contrastive Learning (AFECL) to achieve edgeedge contrast. AFECL depends on no augmentation consisting of two parts. Firstly, we design a novel edge feature generation method, where edge features are computed by embedding concatenation of their connected nodes. Secondly, an edge contrastive learning scheme is developed, where edges connecting the same nodes are defined as positive pairs, and other edges are defined as negative pairs. Experimental results show that compared with recent state-of-the-art GCL methods or even some supervised GNNs, AFECL achieves SOTA performance on link prediction and semi-supervised node classification of extremely scarce labels. The source code is available at https://github.com/YujunLi361/AFECL.

Figures (4)

• Figure 1: The architecture of the proposed AFECL framework. Firstly, the original graph is regarded as a view and fed into the encoder to learn the representation of the nodes. Then, the learned node representations are used to concatenate and generate edge representations. Note that in large graphs, edges need to be sampled first to generate representations behind the sampling. In the end, the model applies edge contrastive loss to maximize the mutual information between positive pairs of representations and minimize the mutual information between negative pairs of representations. The red edges represent both anchors and positives.
• Figure 2: Sensitivity analysis of the hyperparameters $K$, $F^{\prime}$ and $\tau$ on our model.
• Figure 3: t-SNE visualization of representations on benchmark datasets. The rows represent the visualization results on Cora, CiteSeer, and Coauthor-CS, respectively. The columns suggest the results on the original dataset, GCA, GraphACL, and AFECL.
• Figure 4: Experimental results for edge sampling. Orange columns denote the classification accuracy for different numbers of edges when 2 labeled training nodes per class; Blue columns denote the classification accuracy for different numbers of edges when 4 labeled training nodes per class. Specifically, in the picture 6 represents 60,000 edges.