Seeing All From a Few: Nodes Selection Using Graph Pooling for Graph Clustering.

Area of Science:

• Graph theory
• Machine learning
• Data mining

Background:

• Graph clustering methods often struggle with unreliable graph structures containing spurious edges.
• Existing approaches may be negatively impacted by noisy or irrelevant connections within the graph data.

Purpose of the Study:

• To propose a novel dual graph embedding network (DGEN) for node clustering that mitigates the effects of spurious edges.
• To introduce a graph pooling technique, Neighbor Cluster Pooling (NCPool), for selecting informative graph components.

Main Methods:

• Developed a dual graph embedding network (DGEN) with a two-step graph encoder and a graph pooling layer.
• Proposed Neighbor Cluster Pooling (NCPool) to identify informative nodes and edges based on proximity to cluster centers.
• Trained a classifier on selected node clustering results to assign clusters to all nodes.

Main Results:

• The proposed DGEN effectively alleviates the impact of spurious edges on graph clustering.
• Experiments on five benchmark datasets show superior performance compared to state-of-the-art graph clustering algorithms.
• The NCPool mechanism successfully selects informative subsets of nodes and edges.

Conclusions:

• The DGEN framework offers a robust solution for graph clustering in the presence of noisy graph structures.
• Employing graph pooling techniques like NCPool is a promising direction for enhancing node clustering.
• The method demonstrates significant improvements in clustering accuracy and robustness.