Graph explicit pooling for graph-level representation learning.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Representation Learning

Background:

• Graph pooling is essential for hierarchical representation learning in Graph Neural Networks (GNNs).
• Existing methods often lack thorough node impact evaluation and discard potentially useful information from dropped nodes.
• This leads to suboptimal graph coarsening and reduced predictive performance.

Purpose of the Study:

• To address limitations in current graph pooling techniques for GNNs.
• To introduce a novel node selection strategy that considers node relationships and final representation vectors.
• To leverage information from discarded graph segments to enhance learning.

Main Methods:

• Developed Graph explicit Pooling (GrePool) for improved node selection based on node-representation vector relationships.
• Introduced GrePool+ which applies uniform loss on discarded nodes to retain latent information.
• Conducted extensive experiments on 12 benchmark datasets, including Open Graph Benchmark.

Main Results:

• GrePool consistently outperformed 14 baseline graph pooling methods across most datasets.
• GrePool+ further enhanced GrePool's performance without increasing computational overhead.
• The proposed methods demonstrated significant improvements in graph representation learning and classification accuracy.

Conclusions:

• GrePool offers a more effective approach to graph pooling in GNNs by improving node selection and information utilization.
• GrePool+ provides an efficient enhancement, boosting performance by leveraging discarded node information.
• The findings suggest a new direction for developing advanced graph pooling strategies in deep learning.