FlowX: Towards Explainable Graph Neural Networks via Message Flows

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are powerful machine learning models for graph-structured data.
• Current explainability methods for GNNs often focus on nodes, edges, or features, limiting deeper mechanistic understanding.
• Understanding the internal workings of GNNs is crucial for trust and reliable deployment.

Purpose of the Study:

• To develop a novel method for explaining GNNs by analyzing their inherent message flow mechanisms.
• To provide a more natural and effective approach to GNN explainability compared to existing feature-based methods.
• To enhance the interpretability of GNNs for diverse scientific and real-world applications.

Main Methods:

• Proposed FlowX, a novel method to explain GNNs by identifying and quantifying the importance of message flows.
• Utilized Shapley values from cooperative game theory to measure the importance of message flows.
• Developed a flow sampling scheme for efficient computation of Shapley value approximations.
• Introduced an information-controlled learning algorithm to train flow scores for necessary or sufficient explanations.

Main Results:

• Demonstrated that FlowX effectively identifies important message flows within GNNs.
• Experimental results on synthetic and real-world datasets show improved GNN explainability using FlowX.
• The proposed method provides a more intuitive understanding of GNN decision-making processes.

Conclusions:

• Message flows represent a more natural and effective basis for GNN explainability.
• FlowX offers a significant advancement in understanding and interpreting GNN behavior.
• This work paves the way for more transparent and trustworthy GNN models.