Deep attributed graph clustering with self-separation regularization and parameter-free cluster estimation.

Area of Science:

• Graph analysis
• Machine learning
• Data mining

Background:

• Graph neural networks (GNNs) excel at attributed graph clustering by aggregating node features and structure.
• Existing GNN methods often overlook learning cluster-friendly features and require a predefined number of clusters.
• This limits their ability to capture complex cluster characteristics and adapt to varying cluster counts.

Purpose of the Study:

• To develop a deep attributed clustering method that overcomes limitations of existing GNNs.
• To enable learning of cluster-friendly features and automatic determination of the cluster number.
• To improve the performance and flexibility of attributed graph clustering.

Main Methods:

• A jumping graph convolutional auto-encoder is jointly optimized with a self-separation regularizer to learn features that promote dense intra-cluster and sparse inter-cluster structures.
• A softmax auto-encoder is employed for parameter-free estimation of the natural cluster number directly from the data.
• These components are integrated into a deep attributed clustering framework.

Main Results:

• The proposed method successfully learns cluster-friendly features, enabling the discovery of clusters with varying sizes.
• It effectively addresses the limitation of requiring a pre-assigned cluster number by estimating it automatically.
• Extensive experiments demonstrate the superior effectiveness of the proposed model compared to existing approaches.

Conclusions:

• The developed deep attributed clustering method significantly advances the field by learning robust cluster-friendly features and enabling parameter-free cluster number estimation.
• This approach offers a more flexible and effective solution for analyzing complex attributed graphs.
• The findings highlight the potential of self-separated graph neural networks and parameter-free estimation in graph clustering tasks.