Accumulated local effects and graph neural networks for link prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are powerful for link prediction.
• Model-agnostic explanation methods like Accumulated Local Effects (ALE) are crucial for understanding GNNs.
• Directly applying ALE to GNNs is computationally intensive due to complex node interactions.

Purpose of the Study:

• To adapt the Accumulated Local Effects (ALE) method for visualizing node feature influence in GNN-based link prediction.
• To address the computational challenges of applying ALE to GNNs by proposing an approximate method.

Main Methods:

• Investigated the adaptation of Accumulated Local Effects (ALE) for GNNs (Graph Convolutional Networks and Graph Attention Networks).
• Developed and evaluated an approximate method to mitigate the computational cost of ALE in GNNs.
• Analyzed the impact of parameter variations on ALE estimation accuracy for both exact and approximate methods.

Main Results:

• The approximate ALE method significantly improves computational efficiency compared to the exact method.
• The exact ALE method provides more stable explanations, especially with smaller datasets.
• Explanations from the approximate method are comparable to those from the exact method.

Conclusions:

• The approximate ALE method is a viable and efficient approach for explaining GNNs in link prediction tasks.
• Trade-offs between computational efficiency and explanation stability exist, but the approximate method offers a practical solution.
• Further analysis of parameter effects is important for accurate ALE estimation in GNNs.