Graph Multi-Convolution and Attention Pooling for Graph Classification.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Graph classification is crucial for analyzing complex structured data.
• Existing methods struggle with node feature fusion and information loss during pooling.
• Limited attention to multi-hop neighborhood information hinders performance.

Purpose of the Study:

• To propose a novel Graph Multi-Convolution and Attention Pooling (GMCAP) method for enhanced graph classification.
• To address limitations in node feature fusion and information preservation in graph pooling.
• To improve the learning of effective graph-level representations.

Main Methods:

• Developed Graph Multi-Convolution (GMConv) layers for explicit node feature fusion.
• Implemented a weight-based aggregation module to exploit multi-hop neighborhood information.
• Introduced Local information and Global Attention based Pooling (LGAPool) to minimize information loss.

Main Results:

• GMCAP demonstrates superior performance compared to state-of-the-art methods.
• The method effectively fuses node features from diverse perspectives.
• LGAPool successfully reduces information loss during graph pooling.

Conclusions:

• GMCAP offers an effective approach for learning graph-level representations.
• The proposed method enhances graph classification accuracy.
• GMCAP provides a robust solution for analyzing graph-structured data.