cc-Shapley: Measuring Multivariate Feature Importance Needs Causal Context
This addresses the issue of misinterpretations in explainable AI for researchers and practitioners, offering a causal correction to a widely used method, though it is incremental as it modifies existing Shapley values.
The paper tackled the problem of spurious associations in multivariate feature importance using Shapley values, showing that causal knowledge is needed to correct misleading attributions, and proposed cc-Shapley, which theoretically eliminates collider bias and demonstrates nullification or reversal of associations in synthetic and real-world datasets.
Explainable artificial intelligence promises to yield insights into relevant features, thereby enabling humans to examine and scrutinize machine learning models or even facilitating scientific discovery. Considering the widespread technique of Shapley values, we find that purely data-driven operationalization of multivariate feature importance is unsuitable for such purposes. Even for simple problems with two features, spurious associations due to collider bias and suppression arise from considering one feature only in the observational context of the other, which can lead to misinterpretations. Causal knowledge about the data-generating process is required to identify and correct such misleading feature attributions. We propose cc-Shapley (causal context Shapley), an interventional modification of conventional observational Shapley values leveraging knowledge of the data's causal structure, thereby analyzing the relevance of a feature in the causal context of the remaining features. We show theoretically that this eradicates spurious association induced by collider bias. We compare the behavior of Shapley and cc-Shapley values on various, synthetic, and real-world datasets. We observe nullification or reversal of associations compared to univariate feature importance when moving from observational to cc-Shapley.