Work

From Predictive Analytics to Intelligent Intervention: Integrating Causal Machine Learning, Treatment Optimization, and Decision Utility Assessment

Year: 2026
Type: preprint
Abstract: Predictive accuracy does not establish that an intervention is effective. This study proposes a decision-centric artificial intelligence workflow combining predictive modelling, SHAP explainability, heterogeneous treatment-effect estimation, policy learning, and intervention-utility evaluation. Four predictive models (Logistic Regression, Random Forest, XGBoost, and LightGBM) and six causal and uplift estimators (S-Learner, T-Learner, X-Learner, LinearDML, Causal Forest, and Random Forest Two-Model Uplift) were evaluated using a treatment-balanced sample of 500,000 observations from the Criteo benchmark. Binary visit was the primary outcome, while conversion was used descriptively. XGBoost and LightGBM achieved validation AUC-ROC values of 0.94467 and 0.94455, respectively. The Spearman rank correlation between XGBoost predicted visit probabilities and Causal Forest estimated treatment effects was 0.3942 (95% bootstrap CI: 0.3873–0.4020), indicating a moderate association between predictive response and estimated incremental treatment benefit. At a 30% intervention budget, full-test-set policy values were 0.047333 for XGBoost and 0.046827 for Causal Forest, with corresponding net benefits of 0.045833 and 0.045327. In a separate 20,000-observation treatment-stratified evaluation subset, paired-bootstrap comparisons produced a Causal Forest–XGBoost policy-value difference of 0.001200 (95% CI: −0.000903 to 0.003600) and a net-benefit difference of 0.001177 (95% CI: −0.000945 to 0.003562). Neither difference was statistically distinguishable from zero. These findings demonstrate that predictive discrimination, treatment-effect ranking, and intervention utility represent distinct evaluation objectives. The results provide empirical evidence of their complementary roles within the Criteo benchmark but do not establish universal causal-policy superiority or cross-domain generalizability.
Institution: Clark University
Language: English
Open Access status: green
OpenAlex ID: W7236748930
Cites: 30
Cited by:
Related to: 10