Causal Inference
Did the intervention work, and for whom? Causal inference is how you answer that from data that only shows you correlations. My work here covers treatment effect estimation with machine learning and the step that follows it, policy learning, where the estimate decides who receives the intervention. The estimation machinery itself is described on the Generalized Riesz Regression page.
Problems addressed
When a campaign is evaluated, comparing the people who received it against those who did not gives the wrong answer, because the two groups differ in ways that also affect the outcome. Machine learning corrects for that difference well, but the flexibility of the model introduces a bias of its own, and removing it takes care. I have built a single framework for estimating the correction term this requires, released as the Python package genriesz.
Once the effect is estimated, the question becomes which intervention to give to whom. That brings in off-policy evaluation, where the value of a new policy is estimated from logs collected under an old one, and it brings in external validity, where an experimental result has to carry over to a population with a different covariate distribution. Recent work extends this to a Bayesian formulation of policy learning (General Bayesian Policy Learning, UAI 2026) and to retrieval-based action selection.
Research thread
- 2026
- ScoreMatchingRiesz: Score Matching for Debiased Machine Learning and Policy Path Estimation (ICML 2026). Estimates the bias-correction term by score matching and extends it to the whole policy path.
- 2023
- CATE Lasso: Conditional Average Treatment Effect Estimation with High-Dimensional Linear Regression. A Lasso that stays consistent for the conditional average treatment effect in high dimensions.
- 2022
- Learning Causal Models from Conditional Moment Restrictions by Importance Weighting (ICLR 2022). Learning causal models from conditional moment restrictions.
- 2020
- Off-Policy Evaluation and Learning for External Validity under a Covariate Shift (NeurIPS 2020). Evaluating a policy in advance when the target population has shifted.
- 2020
- Off-Policy Evaluation of Bandit Algorithm from Dependent Samples under Batch Update Policy. Policy evaluation from the dependent logs a bandit algorithm produces.
Relation to earlier work
The standard estimation framework for causal inference with machine learning is the double/debiased machine learning of Chernozhukov and coauthors (2018, Double/debiased machine learning for treatment and structural parameters). On the policy side, Athey and Wager gave statistical guarantees for learning a policy from observational data (Policy Learning with Observational Data). My work sits on both lines: it unifies the estimation of the bias-correction term under Bregman divergences, and it extends policy learning toward experimental design and a Bayesian formulation.
Where it applies
Measuring the effect of advertising and promotions, deciding which customers to target, and evaluating a candidate policy against past logs before running it are all direct uses of these methods. For inquiries about effect measurement, from designing the evaluation through to implementing the estimator, write to mkato-csecon@g.ecc.u-tokyo.ac.jp.