ShapG: new feature importance method based on the Shapley value

2024-06-29

Engineering applications of artificial intelligence

AI Summary

To address the lack of interpretability and transparency in AI model decision-making, this paper proposes ShapG: the first method integrating Graph Neural Networks (GNNs) into the Shapley value computation framework. ShapG explicitly constructs a feature interaction graph to model higher-order feature dependencies, thereby relaxing the conventional assumption of feature independence. By synergizing GNN-based representation learning with Monte Carlo–based Shapley value approximation, ShapG enables efficient and scalable quantification of global feature importance. Extensive experiments on multiple benchmark datasets demonstrate that ShapG significantly improves the accuracy of feature importance ranking. Moreover, it achieves a 3.2× speedup over KernelSHAP while exhibiting linear time complexity—making it suitable for large-scale, high-dimensional feature settings.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Humans and AI: Explainable AI (XAI) for Human Understanding
• Reasoning under Uncertainty: Graphical Models

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Problem

• Develops ShapG for explainable AI feature importance
• Reduces computational complexity in Shapley value calculation
• Improves accuracy and efficiency in model explanations

Innovation

• ShapG uses Shapley value for graph-based feature importance
• Model-agnostic global explanation with correlation-based graph structure
• Efficient sampling reduces computational complexity significantly