Manifold-based Shapley explanations for high dimensional correlated features.

Area of Science:

• Artificial Intelligence
• Machine Learning Interpretability
• Network Analysis

Background:

• Explainable Artificial Intelligence (XAI) is crucial for transparent and reliable network decision-making.
• SHapley Additive exPlanations (SHAP) is a common method for interpreting model features but assumes feature independence.
• This independence assumption leads to inaccuracies when dealing with correlated features in high-dimensional data.

Purpose of the Study:

• To introduce Latent SHAP, a novel manifold-based approach to address SHAP's limitations with correlated features.
• To correct misinterpretations in network analysis caused by SHAP's assumption of feature independence.
• To reduce algorithmic complexity for more efficient interpretation of complex networks.

Main Methods:

• Latent SHAP transforms high-dimensional data into low-dimensional manifolds to capture feature correlations.
• Shapley values are computed on the derived data manifold.
• Three gradient-based mapping techniques are employed to transfer Shapley values back to the high-dimensional space.

Main Results:

• Latent SHAP successfully corrects misinterpretations of SHAP for specific network samples.
• The method effectively accounts for feature correlations in high-dimensional data interpretation.
• Demonstrated reduction in algorithmic complexity for complex network interpretation tasks.

Conclusions:

• Latent SHAP offers a robust solution for accurate network interpretation in the presence of feature correlations.
• This manifold-based approach enhances the reliability and transparency of explainable artificial intelligence.
• The developed method provides a more sophisticated tool for understanding complex, high-dimensional network models.