Speaker
Description
Explainable AI (XAI) approaches, most notably Shapley values, have become increasingly popular because they reveal how individual features contribute to a model’s predictions. At the same time, global sensitivity analysis (GSA) techniques, especially Sobol indices, have long been used to quantify how uncertainty in each input (and combinations of inputs) propagates to uncertainty in the model’s output. Prior work (e.g., Owen 2014) highlighted the theoretical connections between Shapley‐based explanations and Sobol‐based sensitivity measures.
FANOVA (functional ANOVA) graphs were created to visualize main‐effect Sobol indices and total interaction terms in a straightforward graphic form, making it easy to see which inputs (and pairs of inputs) drive model behavior. In this work, we apply the same general concept to Shapley values and the recently introduced Shapley interaction indices. By translating complex machine‐learning models into analogous "Shapley graphs", we provide equally intuitive visual representations of both individual feature contributions and feature interactions. Through several real‐world case studies, we show that these Shapley‐based graphs are just as clear and user‐friendly as FANOVA graphs, and we discuss how the two methods compare in terms of interpretability.
References
Owen, A. B. (2014). Sobol’ indices and Shapley value. SIAM/ASA Journal on Uncertainty Quantification, 2, 245–251.
Fruth, J.,Roustant, O. & Kuhnt, S. (2014). Sensitivity Analysis and FANOVA Graphs for Computer Experiments. Journal of Statistical Planning and Inference, 147, 212–223
Muschalik, M., Baniecki, H., Fumagalli, F., Kolpaczki, P., Hammer, B., & Hüllermeier, E. (2024). SHAP-IQ: Shapley Interactions for Machine Learning. In Proceedings of the Thirty-Eighth Conference on Neural Information Processing Systems Datasets and Benchmarks Track
| Classification | Both methodology and application |
|---|---|
| Keywords | XAI, Shapley values, Sobol Indices |