反事实预测是因果推断领域中的一类重要任务。它旨在回答当环境中的特定因素发生变化时，其他因素会如何改变的问题。而在决策问题中，人们关心决策变化对效果产生的影响，并进行不同决策之间的比较。因此，反事实预测对于决策的评估和优化有着非常重大的帮助。传统的因果推断方法基于变量之间的因果关系实现反事实预测，但对数据的产生机制有较强的约束，应用范围受到了限制。机器学习技术虽然在许多复杂预测任务上有出色的表现，但是其基于数据相关性的预测机制使得它在反事实预测问题中表现不佳。因此，本文结合因果推断和机器学习的技术和优势，考察复杂现实场景下的决策问题，开展了基于反事实预测的决策评估和优化方法研究。具体地，本文分析了在现实决策场景下可能遇到的挑战，包括信息缺失、复杂决策变量以及样本量有限等，提出了相应的反事实预测方法，并通过实验证明了将因果推断和机器学习相结合的反事实预测在决策评估和优化任务上具有重大价值。本文的研究内容和创新成果包括以下部分：? 复杂决策变量下的反事实决策评估：传统的干预效果反事实预测研究，通常考虑二值等简单类型干预的场景。针对复杂的高维干预场景下的干预效果预测问题，本文提出了变分样本重加权方法，学习高维干预的低维隐表征结构来降低问题复杂度，显著地提升了的干预效果反事实预测的准确度。? 信息缺失下的反事实决策评估：在现实的决策评估问题中，人们往往会遇到信息缺失的问题。本文重点考虑混淆变量缺失和原始策略信息缺失两类场景。对于混淆变量缺失下的干预效果预测问题，本文提出了基于事实观察的异质性学习方法，还原隐混淆变量，显著地降低了干预效果反事实预测的误差。对于原始策略信息缺失下的策略评估问题，本文利用了因果推断中的变量平衡技术，从而避开了对原始策略信息的依赖。进一步地，本文将待评估策略的信息引入到权重学习中，使得策略效用值的评估结果更加精准。? 样本量有限下的反事实决策优化：传统的干预效果预测模型旨在优化模型对不同干预的效果预测总误差。这类优化目标在干预空间较大、样本量有限的场景下，与决策优化的目标并不一致。本文提出了效果导向的样本重加权方法，使得反事实预测模型在训练过程中更加关注于效果较好的干预区域，最终使得基于反事实预测模型所学到的策略的效用值得到了显著的提升。

Counterfactual prediction is a crucial task in the field of causal inference. It aims to answer the question of how other factors would changed when certain factors in the environment were altered. In decision-making problems, people are concerned about the impact of decision changes on outcome and engage in comparing different decisions. Therefore, counterfactual prediction plays a significant role and provides substantial assistance in the decisions evaluation and optimization. Conventional causal inference methods achieve counterfactual prediction based on the causal relationships between variables, but have strong constraints on the data generation process, which brings restriction to their applicability. Though machine learning technologies exhibits outstanding performance in many complex prediction tasks, the prediction mechanism based on data correlations makes it perform poorly in the counterfactual prediction tasks. Therefore, this paper combines the techniques and advantages of causal inference and machine learning, investigates the decision-making problem in complex real-world scenarios and conduct research on decision evaluation and optimization based on counterfactual prediction. Specifically, this paper analyzes the potential challenges that may arise in the real decision-making scenarios, including missing information, complex decision variables and limited samples, and proposes the corresponding methods of counterfactual prediction. By experiments, this paper verifies that the counterfactual prediction integrating both causal inference and machine learning is of great value for decision evaluation and optimization. The main research contents and novel contributions are summarized as following:Counterfactual decision evaluation with complex decision variables: Conventional research on counterfactual prediction of treatment outcome usually considers the scenarios involving simple types of treatments, such as binary treatments. For the counterfactual prediction problem of high-dimensional treatments, this paper proposes variational sample re-weighting method which learns the low-dimensional latent representation of high-dimensional treatments to reduce the complexity of the problem, and significantly improves the accuracy of counterfactual prediction on treatment outcome. Counterfactual decision evaluation with missing information: In the real decision evaluation scenarios, people usually encounter missing information problems. This paper focuses on two types of scenarios: missing confounders and missing information of original policies. For the problem of treatment outcome prediction with missing confounders, this paper proposes factual observation based heterogeneity learning method to recover the hidden confounders, and significantly reduces the counterfactual prediction error of treatment outcome. For the problem of policy evaluation with missing information of original policies, this paper introduces the covariate balancing method of causal inference, and therefore bypasses the reliance on the information of original policies. Furthermore, this paper incorporates the information of target policy into learning sample weights, and makes the evaluation result more accurate. Counterfactual decision optimization with limited data:Conventional counterfactual prediction model aims to minimize the total outcome prediction error of different treatments. In the scenarios with large treatment space and limited data, this optimization target is inconsistent with decision optimization. This paper proposes outcome-oriented sample re-weighting method to make the model focus more on the treatment region with better outcome during training, and finally makes the policy learned from counterfactual prediction model have significantly improved utility.