Embedding Graph Auto-Encoder for Graph Clustering.

Area of Science:

• Machine Learning
• Graph Theory
• Data Mining

Background:

• Graph clustering partitions nodes into groups using unsupervised methods.
• Graph Auto-Encoder (GAE) models, based on semi-supervised Graph Convolutional Networks (GCN), show promise but often ignore representation orthogonality or decouple training.
• Existing methods fail to fully utilize the properties of GAE representations for optimal clustering.

Purpose of the Study:

• To develop a novel Graph Auto-Encoder (GAE) based model for graph clustering.
• To address limitations of existing methods by integrating representation learning and clustering.
• To leverage theoretical properties of relaxed k-means for improved graph clustering.

Main Methods:

• Developed Embedding Graph Auto-Encoder (EGAE), a GAE-based model for graph clustering.
• Integrated relaxed k-means theory with GAE for simultaneous learning of representations and clustering.
• Designed EGAE with an encoder and dual decoders, where relaxed k-means acts as a decoder.

Main Results:

• EGAE demonstrates superior performance in graph clustering compared to existing methods.
• The learned representations in EGAE are explainable and applicable to other tasks.
• Simultaneous learning of GAE and relaxed k-means leads to effective deep feature generation for clustering.

Conclusions:

• EGAE offers a theoretically grounded and effective approach to unsupervised graph clustering.
• The model's ability to jointly optimize representation learning and clustering enhances performance.
• EGAE's explainable representations broaden its utility beyond clustering.