Boosting Pseudo-Labeling With Curriculum Self-Reflection for Attributed Graph Clustering.

Area of Science:

• Machine Learning
• Graph Theory
• Data Mining

Background:

• Attributed graph clustering partitions nodes into groups using unsupervised learning.
• Current methods often introduce bias and label noise through pretext tasks.
• Predictive methods show promise but face limitations.

Purpose of the Study:

• To propose a novel self-supervised learning method for attributed graph clustering.
• To address limitations of existing predictive methods, specifically auxiliary task bias and label noise.
• To improve the accuracy and reliability of graph clustering.

Main Methods:

• Introduced Pseudo-Labeling with Curriculum Self-Reflection (PLCSR).
• Employed a self-auxiliary encoder using exponential moving average (EMA) for confident pseudo-label generation.
• Implemented a curriculum selection strategy with dynamic thresholds for accurate node processing.

Main Results:

• PLCSR significantly outperforms state-of-the-art predictive methods.
• Achieved over 6% improvement in clustering accuracy compared to CDRS.
• Demonstrated effective handling of node uncertainty through abstention.

Conclusions:

• PLCSR offers a robust and accurate approach to attributed graph clustering.
• The method mitigates bias and label noise inherent in prior techniques.
• PLCSR advances the field of self-supervised graph representation learning.