A comparative study of methods for estimating model-agnostic Shapley value explanations

• Published in: Data mining and knowledge discovery Vol. 38; no. 4; pp. 1782 - 1829
• Main Authors: Olsen, Lars Henry Berge, Glad, Ingrid Kristine, Jullum, Martin, Aas, Kjersti
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.07.2024; Springer Nature B.V
• Subjects: Artificial Intelligence; Chemistry and Earth Sciences; Comparative studies; Computer Science; Data Mining and Knowledge Discovery; Estimation; Game theory; Information Storage and Retrieval; Machine learning; Methods; Monte Carlo simulation; Physics; Prediction models; Predictions; Regression models; Statistics for Engineering; Explainable artificial intelligence; Shapley values; Feature importance; Feature dependence; Prediction explanation; Model-agnostic explanation
• ISSN: 1384-5810; 1573-756X
• DOI: 10.1007/s10618-024-01016-z

Shapley values originated in cooperative game theory but are extensively used today as a model-agnostic explanation framework to explain predictions made by complex machine learning models in the industry and academia. There are several algorithmic approaches for computing different versions of Shapley value explanations. Here, we consider Shapley values incorporating feature dependencies, referred to as conditional Shapley values, for predictive models fitted to tabular data. Estimating precise conditional Shapley values is difficult as they require the estimation of non-trivial conditional expectations. In this article, we develop new methods, extend earlier proposed approaches, and systematize the new refined and existing methods into different method classes for comparison and evaluation. The method classes use either Monte Carlo integration or regression to model the conditional expectations. We conduct extensive simulation studies to evaluate how precisely the different method classes estimate the conditional expectations, and thereby the conditional Shapley values, for different setups. We also apply the methods to several real-world data experiments and provide recommendations for when to use the different method classes and approaches. Roughly speaking, we recommend using parametric methods when we can specify the data distribution almost correctly, as they generally produce the most accurate Shapley value explanations. When the distribution is unknown, both generative methods and regression models with a similar form as the underlying predictive model are good and stable options. Regression-based methods are often slow to train but quickly produce the Shapley value explanations once trained. The vice versa is true for Monte Carlo-based methods, making the different methods appropriate in different practical situations.