Smoothness Sensor: Adaptive Smoothness-Transition Graph Convolutions for Attributed Graph Clustering.

Area of Science:

• Graph Machine Learning
• Data Mining
• Network Science

Background:

• Clustering groups similar objects; attributed graph clustering integrates node features and structure.
• Graph convolutional networks (GCNs) effectively combine attributes and structure for clustering.
• Oversmoothing in GCNs degrades node representations, hindering clustering performance.

Purpose of the Study:

• To develop novel methods to prevent oversmoothing in GCNs for attributed graph clustering.
• To introduce adaptive smoothness-transition graph convolutions and fine-grained nodewise smoothness assessment.
• To enhance clustering accuracy by preserving distinct node representations.

Main Methods:

• Proposed a smoothness sensor with adaptive smoothness-transition graph convolutions to prevent oversmoothing.
• Introduced a fine-grained nodewise-level smoothness assessment based on neighborhood conditions.
• Designed a self-supervision criterion focusing on intra-cluster tightness and inter-cluster separation.

Main Results:

• The proposed methods significantly outperformed 13 state-of-the-art baselines across five benchmark datasets.
• Demonstrated improved clustering performance in terms of various evaluation metrics.
• Experimental analysis confirmed the effectiveness and efficiency of the proposed approaches.

Conclusions:

• The developed smoothness sensor and nodewise assessment effectively mitigate oversmoothing in GCNs.
• The novel approach enhances attributed graph clustering by maintaining discriminative node representations.
• This work offers a promising direction for improving GCN-based clustering techniques.