Structure-missing graph-level clustering network.

Area of Science:

• Graph representation learning
• Machine learning
• Data mining

Background:

• Existing graph-level clustering methods assume complete graph structures, failing to account for missing relationships common in real-world data.
• Missing relationships lead to structural information distortion, significantly degrading clustering performance.
• The problem of graph-level clustering with missing relations is novel and underexplored.

Purpose of the Study:

• To propose a novel method, Structure-Missing Graph-Level Clustering Network (SMGCN), designed to handle missing relationships in graph-level clustering.
• To improve the accuracy and robustness of graph clustering by addressing structural information distortion.
• To introduce a new benchmark task for graph clustering research focusing on incomplete graph data.

Main Methods:

• Structure augmentation using a low-rank matrix completion module (LR-SEA) to reconstruct missing relationships.
• An Anchor Positioning Mechanism utilizing the Hungarian algorithm for cluster matching and anchor identification.
• Joint Contrastive Optimization to align graph embeddings with identified anchors, forcing similar clusters to converge.

Main Results:

• The proposed SMGCN method demonstrates superior performance compared to state-of-the-art methods.
• Experiments on five benchmark datasets validate the effectiveness of SMGCN in handling missing relationships.
• The method successfully mitigates structural information distortion caused by incomplete graph data.

Conclusions:

• SMGCN effectively addresses the challenge of graph-level clustering with missing relationships.
• The proposed approach enhances graph representation learning by reconstructing and utilizing structural information.
• This work establishes a new direction for graph clustering research on incomplete datasets.