Interpretable Uplift Modelling: Conditional Average Treatment Effect Attribution, Subgroup Rules and Constrained Recourse

Uplift models estimate how treatment effects vary across individuals, but a prediction alone does not explain feature contributions, define usable subgroups, or suggest alternative feature states. We combine these three outputs through one fitted conditional average treatment effect (CATE) model. Treatment Effect Shapley Values (TESV) applies existing interventional Shapley attribution directly to this model. Stable Causal Rule Learning (SCRL) produces concise subgroup descriptions using empirical treatment–control support checks and a bootstrap‐based score. Constrained feature‐space recourse searches for higher predicted CATE under permitted coordinate changes. Across four benchmarks, we evaluate these outputs against known effect modifiers, the fitted model, and held‐out outcome contrasts, keeping these reference targets distinct. Attribution reflects dependencies learned by the model; marginal replacement does not remove learned proxy dependence. Rule quality varies with the dataset and evaluation target, and the Hillstrom randomised evaluation does not establish higher per‐contact efficacy than the model's ranking. The framework therefore supports model inspection, while its outputs require further evidence before guiding interventions. Empirical support checks do not establish causal identification, and recourse targets are neither validated actions nor guaranteed jointly plausible or reachable states.