A SCALABLE CAUSAL MACHINE LEARNING FRAMEWORK FOR DATA-DRIVEN DECISION OPTIMIZATION
Keywords:
Causal machine learning, Heterogeneous treatment effects, Uplift modeling, Doubly robust learning, Stacking, Budget-constrained optimization, Policy learning, ScalabilityAbstract
Allocating costly interventions such as incentives, outreach messages or preventive programs to the individuals who benefit most is a central task of data-driven decision making. It requires counterfactual predictions for several competing treatments, a principled way to choose among the many available causal learners, and an optimizer that respects budgets at the scale of millions of units. This paper presents SCOPE (Scalable Causal Orthogonal Policy Ensemble), an end-to-end framework that couples orthogonal machine learning with budgeted multi-treatment optimization. SCOPE builds a cross-fitted nuisance cache with histogram gradient boosting, trains a leakage-free library of S-, T-, X-, R- and DR-learners, aggregates them by arm-wise orthogonal stacking with non-negative, free-scale weights fitted on a doubly robust loss, and solves the resulting multiple-choice knapsack problem with a dual-price bisection whose gap to the linear relaxation is bounded by the gain of a single unit. Every stage runs in time linear in the sample size. On 50 IHDP replications SCOPE attains the lowest √PEHE (2.12 versus 2.14) and the highest policy value at every budget. On a field experiment with 305,866 voters and four arms it raises the turnout gain of a 10% mailing budget by 8.9% over the best single learner and by 17.8% over random targeting, and on a confounded multi-treatment benchmark it reaches 92.4–95.2% of the oracle value. The dual allocator matches the linear-programming relaxation within 0.010% while running about 83 times faster.References
[1] Künzel S R, Sekhon J S, Bickel P J, et al. Metalearners for estimating heterogeneous treatment effects using machine learning. Proceedings of the National Academy of Sciences of the United States of America, 2019, 116(10): 4156-4165.
[2] Nie X, Wager S. Quasi-oracle estimation of heterogeneous treatment effects. Biometrika, 2021, 108(2): 299-319.
[3] Kennedy E H. Towards optimal doubly robust estimation of heterogeneous causal effects. Electronic Journal of Statistics, 2023, 17(2): 3008-3049.
[4] Chernozhukov V, Chetverikov D, Demirer M, et al. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 2018, 21(1): C1-C68.
[5] Foster D J, Syrgkanis V. Orthogonal statistical learning. The Annals of Statistics, 2023, 51(3): 879-908.
[6] Wager S, Athey S. Estimation and inference of heterogeneous treatment effects using random forests. Journal of the American Statistical Association, 2018, 113(523): 1228-1242.
[7] Athey S, Tibshirani J, Wager S. Generalized random forests. The Annals of Statistics, 2019, 47(2): 1148-1178.
[8] Shalit U, Johansson F D, Sontag D. Estimating individual treatment effect: Generalization bounds and algorithms. In: Proceedings of the 34th International Conference on Machine Learning (ICML), PMLR 70, 2017: 3076-3085.
[9] Shi C, Blei D M, Veitch V. Adapting neural networks for the estimation of treatment effects. In: Advances in Neural Information Processing Systems 32 (NeurIPS), 2019.
[10] Curth A, van der Schaar M. In search of insights, not magic bullets: Towards demystification of the model selection dilemma in heterogeneous treatment effect estimation. In: Proceedings of the 40th International Conference on Machine Learning (ICML), PMLR 202, 2023: 6623-6642.
[11] Fernández-Loría C, Provost F. Causal decision making and causal effect estimation are not the same… and why it matters. INFORMS Journal on Data Science, 2022, 1(1): 4-16.
[12] Sinha P, Zoltners A A. The multiple-choice knapsack problem. Operations Research, 1979, 27(3): 503-515.
[13] Zhao Z, Harinen T. Uplift modeling for multiple treatments with cost optimization. In: Proceedings of the IEEE International Conference on Data Science and Advanced Analytics (DSAA), 2019.
[14] Ai M, Li B, Gong H, et al. LBCF: A large-scale budget-constrained causal forest algorithm. In: Proceedings of the ACM Web Conference (WWW), 2022: 2310-2319.
[15] Zhou H, Li S, Jiang G, et al. Direct heterogeneous causal learning for resource allocation problems in marketing. In: Proceedings of the AAAI Conference on Artificial Intelligence, 2023, 37(4): 5446-5454.
[16] Hill J L. Bayesian nonparametric modeling for causal inference. Journal of Computational and Graphical Statistics, 2011, 20(1): 217-240.
[17] Gutierrez P, Gérardy J Y. Causal inference and uplift modelling: A review of the literature. In: Proceedings of the International Conference on Predictive Applications and APIs, PMLR 67, 2017: 1-13.
[18] Alaa A, Ahmad Z, van der Laan M. Conformal meta-learners for predictive inference of individual treatment effects. In: Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.
[19] Kitagawa T, Tetenov A. Who should be treated? Empirical welfare maximization methods for treatment choice. Econometrica, 2018, 86(2): 591-616.
[20] Dudík M, Langford J, Li L. Doubly robust policy evaluation and learning. In: Proceedings of the 28th International Conference on Machine Learning (ICML), 2011: 1097-1104.
[21] Athey S, Wager S. Policy learning with observational data. Econometrica, 2021, 89(1): 133-161.
[22] Zhou Z, Athey S, Wager S. Offline multi-action policy learning: Generalization and optimization. Operations Research, 2023, 71(1): 148-183.
[23] Bertsimas D, Kallus N. From predictive to prescriptive analytics. Management Science, 2020, 66(3): 1025-1044.
[24] Elmachtoub A N, Grigas P. Smart ‘predict, then optimize’. Management Science, 2022, 68(1): 9-26.
[25] Mandi J, Kotary J, Berden S, et al. Decision-focused learning: Foundations, state of the art, benchmark and future opportunities. arXiv:2307.13565, 2023.
[26] Breiman L. Stacked regressions. Machine Learning, 1996, 24(1): 49-64.
[27] van der Laan M J, Polley E C, Hubbard A E. Super learner. Statistical Applications in Genetics and Molecular Biology, 2007, 6(1): 25.
[28] Ke G, Meng Q, Finley T, et al. LightGBM: A highly efficient gradient boosting decision tree. In: Advances in Neural Information Processing Systems 30 (NIPS), 2017: 3146-3154.
[29] Gerber A S, Green D P, Larimer C W. Social pressure and voter turnout: Evidence from a large-scale field experiment. American Political Science Review, 2008, 102(1): 33-48.