Hierarchical Pooling in Graph Neural Networks to Enhance Classification Performance in Large Datasets

Area of Science:

• Graph Neural Networks
• Deep Learning
• Machine Learning

Background:

• Deep learning, particularly graph neural networks (GNNs), excels at node classification and prediction by learning node embeddings.
• Differentiable graph pooling (DiffPool) learns cluster assignments for nodes in GNNs but faces challenges with parameter complexity and control.
• End-to-end models in GNNs can suffer from a large number of parameters, potentially leading to redundancy and difficult learning control.

Purpose of the Study:

• To introduce FPool, an advancement over DiffPool for graph representation learning.
• To improve the control and efficiency of deep graph neural network learning processes.
• To enhance data classification and prediction performance on sensor datasets.

Main Methods:

• FPool builds upon the DiffPool framework by modifying the pooling mechanism applied to node representations.
• The proposed method refines the differentiable soft cluster assignment technique for improved node mapping.
• Evaluation involved applying FPool to popular, publicly available sensor datasets.

Main Results:

• FPool demonstrated superior classification and prediction performance compared to existing methods.
• The FPool approach achieved a significant reduction in training time versus the standard DiffPool.
• Experimental results validate the effectiveness of FPool on sensor data.

Conclusions:

• FPool offers an effective improvement over DiffPool for node classification tasks in deep graph neural networks.
• The method enhances performance and training efficiency, addressing limitations of previous approaches.
• FPool shows promise for applications requiring efficient and accurate graph representation learning.