Understanding Pooling in Graph Neural Networks.

Area of Science:

• Graph Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Graph machine learning models often use pooling operators to reduce graph size.
• Existing pooling operators are diverse and lack a unified theoretical framework.

Purpose of the Study:

• To present an operational framework for unifying diverse graph pooling operators.
• To introduce a taxonomy for categorizing pooling operators based on their characteristics.
• To propose evaluation criteria for assessing pooling operator performance.

Main Methods:

• Developed the Selection, Reduction, and Connection (SRC) framework to describe pooling operators.
• Created a taxonomy of pooling operators based on the SRC model.
• Defined three criteria for evaluating pooling operator performance.

Main Results:

• The SRC framework successfully unifies existing pooling operator literature.
• The proposed taxonomy provides a structured way to understand different pooling approaches.
• Performance evaluation revealed distinct behaviors of various operators across different tasks.

Conclusions:

• The SRC framework offers a standardized approach to understanding and developing graph pooling operators.
• The taxonomy and evaluation criteria facilitate the selection of appropriate operators for specific graph machine learning tasks.
• This work advances the field by providing a unified perspective on graph pooling techniques.