GSSCL: A framework for Graph Self-Supervised Curriculum Learning based on clustering label smoothing.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Graph self-supervised learning (GSSL) leverages pretext tasks for unlabeled graph data.
• Existing GSSL methods often use clustering labels, which can introduce noise and lead to overfitting.
• This noise can reduce model performance and generalizability.

Purpose of the Study:

• To propose a novel framework, Graph Self-Supervised Curriculum Learning (GSSCL), to address limitations in existing GSSL methods.
• To improve the generalizability and reduce overfitting in GSSL by employing clustering label smoothing.
• To enhance the reliability of self-supervised signals in graph learning.

Main Methods:

• GSSCL employs a curriculum learning strategy, ordering clusters from easy to difficult.
• It utilizes the Silhouette Coefficient to assess node clustering confidence scores.
• Pseudo-label smoothing is applied to K-nearest neighbor graphs based on feature similarity to handle graph heterophily and noisy links.

Main Results:

• The proposed GSSCL framework demonstrates superior performance across diverse graph benchmarks.
• It achieves comparable results to state-of-the-art methods in semi-supervised node classification.
• The framework shows strong performance in graph clustering tasks.

Conclusions:

• GSSCL effectively reduces reliance on precise clustering, enhancing model generalizability.
• The method successfully mitigates issues arising from noisy clustering labels in GSSL.
• GSSCL offers a robust approach for learning from unlabeled graph data, particularly in complex or heterophilous graph structures.