On exploring node-feature and graph-structure diversities for node drop graph pooling.

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning
• Data Mining

Background:

• Graph pooling is crucial for learning graph-level representations in GNNs.
• Node drop pooling is a leading technique, but often neglects node feature diversity and graph structures.
• This leads to suboptimal performance in graph classification tasks.

Purpose of the Study:

• To introduce a novel plug-and-play scoring scheme, Multidimensional score space (MID), to improve node drop pooling in GNNs.
• To address the limitations of existing methods in capturing node feature diversity and diverse graph structures.
• To enhance the quality of graph-level representations learned by GNNs.

Main Methods:

• Developed MID, a scoring scheme comprising a Multidimensional score space with flIpscore and Dropscore operations.
• flIpscore promotes the preservation of distinct node features.
• Dropscore encourages consideration of a range of graph structures beyond local ones.

Main Results:

• MID was integrated with existing node drop pooling methods (TopKPool, SAGPool, GSAPool, ASAP).
• Significant average performance improvement of approximately 2.8% was observed across 17 real-world graph classification datasets.
• The proposed method demonstrated enhanced graph-level representation learning capabilities.

Conclusions:

• The MID scoring scheme effectively improves node drop pooling methods for GNNs.
• MID enhances graph classification performance by better preserving node feature diversity and graph structures.
• This approach offers a valuable contribution to the field of graph representation learning.