Kun Kuang

Heterogeneous treatment effect (HTE) estimation is vital for understanding how treatment effects vary across individuals or subgroups. Recently, numerous HTE models, such as balancing methods and instrumental variable (IV) methods, have been developed to eliminate confounding bias from both observed variables and unmeasured confounders. However, many of these models overlook distribution shifts across populations, limiting their applicability to in-distribution (ID) scenarios, where the sample distribution matches that of the training dataset. In real-world applications where population distributions are subject to continuous change, there is an urgent need to develop methods for stable HTE estimation across out-of-distribution (OOD) populations. As pioneers in resolving this problem, we propose a novel Stable Balanced Representation Learning with Hierarchical-Attention Paradigm and Instrumental Regularizer (SBRL-HAPI) framework, which consists of four key components: 1) Instrumental Regularizer (IvR) for mitigating unmeasured confounding, 2) Balancing Regularizer (BR) for eliminating confounding bias induced by observed variables, 3) Independence Regularizer (IR) for addressing the distribution shift and 4) Hierarchical-Attention Paradigm (HAP) for coordination between balance and independence. In this way, SBRL-HAPI regresses counterfactual outcomes using ID data, while ensuring the resulting HTE estimation can be successfully generalized to out-of-distribution scenarios, thereby enhancing the model's applicability in real-world settings. Extensive experiments on synthetic and semi-synthetic datasets demonstrate the effectiveness of all four modules including IvR in achieving stable HTE estimation. Furthermore, large-scale industrial experiments on the Kuaishou dataset confirm the OOD robustness of BR, IR, and HAP under realistic distribution shifts.

In many data-driven decision-making scenarios, it is crucial to accurately estimate causal effects on time-to-event outcomes, e.g., time-to-purchase. However, these outcomes are often right-censored: due to events occurring beyond the observation period or loss to follow-up, the observed durations for censored samples are the last follow-up times rather than the actual event times. Critically, this censoring is usually informative : the censoring mechanism is intrinsically linked to the event time, which may be partially missing. This dependency violates the non-informative censoring assumption in conventional survival analysis. Consequently, the latent information in censored data is non-ignorable for causal effect estimation, inducing a censoring bias that remains unresolved by current methodologies. To overcome these challenges, we establish a theoretical framework for identifying causal effects on time-to-event outcomes under informative censoring. Based on this, we propose Multiple Imputation with Shadow Calibration (MISC), a principled framework leveraging shadow variables --- fully observed covariates correlated with event time but conditionally independent of the censoring mechanism --- to enable unbiased estimation. Specifically, MISC (1) automatically learns valid shadow representations from observed covariates, (2) utilizes these representations to identify the conditional distribution of event times to impute censored units, and (3) estimates causal effects using the imputed data. Experimental results on semi-synthetic benchmarks and real-world datasets from medical scenarios and a large-scale e-commerce platform show that MISC achieves state-of-the-art performance in causal effect estimation and counterfactual prediction. Codes are available at: https://github.com/suikateiou/MISC.

This paper studies the problem of Continuous-Time Counterfactual Quantile Learning (CT-CQL) for risk-sensitive policy optimization. In many real-world applications such as patient blood pressure monitoring, financial market analysis, and autonomous driving, data is high-frequency and continuously evolving. However, most existing causal inference methods focus on expectation-based or discrete-time counterfactual reasoning, which fail to capture fine-grained temporal dynamics. As a result, policies optimized under these frameworks may overlook critical risks—e.g., a treatment policy with good average outcomes may still expose patients to life-threatening episodes. To overcome these limitations, we propose CT-CQL, a framework built upon a novel identification theory and featuring three key components: (1) modeling full counterfactual outcome distributions via Stochastic Differential Equations (SDEs) governed by the Fokker–Planck Equation (FPE); (2) enhancing robustness through a minimax objective that minimizes FPE residuals under adversarial perturbations; and (3) mitigating confounding bias using a double-robust AIPW loss. CT-CQL enables robust policy optimization by identifying optimal intervention strategies that maximize expected utility while adhering to real-world budget and safety constraints. Experiments on widely-used benchmarks and the real-world MIMIC-III dataset demonstrate the validation and superiority of the proposed method. The project is available at: https://github.com/Eliza-YiHe/CT-CQL/

Causal inference plays an important role in explanatory analysis and decision-making across a wide range of fields, including statistics, marketing, healthcare, and education. Its core objective is to estimate treatment effects and inform intervention policies. Most existing work focuses on the binary treatment setting, where each unit is assigned to either treatment or control. In practice, however, treatments are often more complex, encompassing multi-valued, continuous, or bundle interventions. We refer to such settings as complex treatments. In this article, we provide a systematic and comprehensive survey of causal inference methods for complex treatments. We first revisit the problem formulation, core assumptions, and their possible variations under different settings. We sequentially review the representative methods for multi-valued, continuous, and bundle treatments. Within each setting, we organize the methods into two broad categories: those that rely on the unconfoundedness assumption and those that address violations of this assumption. We further discuss the intrinsic relationships among these methods and the assumption verification. Finally, we summarize available benchmark datasets and open-source codes, and outline several directions for future research.