shapemonotone

Shapemonotone is a term used in explainable AI to describe approaches that integrate monotonicity constraints into SHAP-based explanations. It refers to methods that aim to guarantee that the attribution of a feature to a prediction behaves monotonically with respect to the feature, for features known to have a monotone relationship with the target.

Background and motivation: SHAP provides local explanations of model predictions by assigning Shapley values to features, reflecting their contribution to a specific outcome. In some domains, practitioners know that increasing a feature should non-decreasingly affect the model’s prediction (for example, higher credit score should not reduce risk). Shapemonotone approaches attempt to align SHAP explanations with this domain knowledge by enforcing monotone behavior in the attribution or in the underlying model.

Methods: Shapemonotone can be realized through several strategies. One route is to train or calibrate models with monotone constraints, such as monotone gradient boosting or monotone neural networks, so that the model itself respects the expected monotonic relationships. Another route is to post-process SHAP values or modify the coalition evaluation to enforce monotonicity across samples or feature values, potentially using optimization with linear inequality constraints or isotonic regression. The goal is to preserve the additive interpretability of SHAP while eliminating explanations that violate known monotonic tendencies.

Applications and limitations: The approach is particularly relevant in regulated or high-stakes domains like finance, healthcare, and risk assessment, where monotone relationships are expected or mandated. Limitations include potential reductions in fidelity to the original model, added computational complexity, and the need for reliable prior knowledge about which features are monotone.

See also: SHAP, Shapley values, monotone constraints, interpretable machine learning, explainable AI.