A Unified Graph Clustering Network
Renda Han, Kuntharrgyal Khysru, Wenxin Zhang, Ronghao Fu, Z. Zhang, Kaiming Wang
Problems Identified (2)
Isolated node-graph clustering: Existing graph clustering methods often handle node-level and graph-level clustering separately, wasting resources and limiting mutual knowledge transfer and performance improvement.
Isolated node-graph clustering: Existing graph clustering methods often handle node-level and graph-level clustering separately, wasting resources and limiting mutual knowledge transfer and performance improvement.
Proposed Solutions (5)
Unified Graph Clustering Network: UGCN collaboratively addresses node-level and graph-level clustering using both local and global graph information.
Dual-branch joint projector: The method uses a dual-branch projector for joint node-level and graph-level learning, combining node prototypes with subgraph and graph representations.
Bidirectional contrastive alignment: The two branches are aligned with joint contrastive objectives so node-level prototypes guide graph clustering and graph-level pseudo-labels improve node clustering.
Unified Graph Clustering Network: UGCN collaboratively addresses node-level and graph-level clustering using both local and global graph information.
Dual-branch joint projector: The method uses a dual-branch projector for joint node-level and graph-level learning, combining node prototypes with subgraph and graph representations.
Results (2)
State-of-the-art clustering performance:
State-of-the-art clustering performance:
Research Domain
Graph neural networks and graph clustering