Affinity Uncertainty-Based Hard Negative Mining in Graph Contrastive Learning.

Area of Science:

• Graph Machine Learning
• Self-Supervised Learning
• Representation Learning

Background:

• Hard negative mining improves contrastive learning (CL), including graph CL (GCL).
• Existing methods struggle with graph data due to oversmoothing and non-i.i.d. issues, leading to false negatives.
• Current hardness-aware CL methods often fail to correctly identify hard negatives in graph data.

Purpose of the Study:

• To propose a novel approach for mining hard negatives in GCL by utilizing collective affinity information.
• To enhance existing GCL methods by incorporating uncertainty information into loss functions.
• To improve the discriminative power of graph representations.

Main Methods:

• A discriminative model is built on collective affinity information (pairwise affinities between anchor and negative instances).
• Model confidence/uncertainty regarding negative instance affinity determines hardness weights.
• Uncertainty information is integrated into GCL loss functions as a weighting term.

Main Results:

• The proposed approach consistently enhances state-of-the-art GCL methods on ten graph datasets.
• Significant improvements were observed in both graph and node classification tasks.
• The method notably boosts the robustness of GCL against adversarial attacks.

Conclusions:

• The novel approach effectively addresses the challenge of hard negative mining in GCL.
• Incorporating uncertainty-based hardness weights leads to improved GCL performance and robustness.
• The enhanced GCL loss is theoretically linked to a triplet loss with an adaptive margin.