This paper introduces a new meta-algorithm called the X-learner for estimating heterogeneous treatment effects (HTE) using machine learning. The X-learner builds on the T-learner and S-learner, which are existing meta-algorithms for estimating the Conditional Average Treatment Effect (CATE). The X-learner is designed to exploit structural properties of the CATE function and perform well when one treatment group is much larger than the other. It is provably efficient under certain conditions, such as when the CATE is linear and the response functions are Lipschitz continuous. The X-learner is implemented using Random Forests (RF) and Bayesian Additive Regression Trees (BART) as base learners. The paper presents simulation studies and field experiments to evaluate the performance of the X-learner and other meta-learners. In the first experiment, the effect of a mailer on voter turnout is estimated, and in the second, the effect of door-to-door conversations on prejudice against gender-nonconforming individuals is measured. The results show that the X-learner can effectively target treatment regimes and shed light on underlying mechanisms. The X-learner outperforms the T-learner and S-learner in many cases, particularly when the CATE function has a simpler structure or when one treatment group is much larger than the other. The paper also discusses the theoretical properties of the X-learner, including its convergence rates and performance under different assumptions. It shows that the X-learner can achieve faster convergence rates when the CATE function is smoother than the response functions. The X-learner is particularly effective when the number of control units is much larger than the number of treated units, as this allows for more accurate estimation of the CATE function. The paper concludes that the X-learner is a valuable tool for estimating HTE in a wide range of applications, and it provides a software package that implements the methods described.This paper introduces a new meta-algorithm called the X-learner for estimating heterogeneous treatment effects (HTE) using machine learning. The X-learner builds on the T-learner and S-learner, which are existing meta-algorithms for estimating the Conditional Average Treatment Effect (CATE). The X-learner is designed to exploit structural properties of the CATE function and perform well when one treatment group is much larger than the other. It is provably efficient under certain conditions, such as when the CATE is linear and the response functions are Lipschitz continuous. The X-learner is implemented using Random Forests (RF) and Bayesian Additive Regression Trees (BART) as base learners. The paper presents simulation studies and field experiments to evaluate the performance of the X-learner and other meta-learners. In the first experiment, the effect of a mailer on voter turnout is estimated, and in the second, the effect of door-to-door conversations on prejudice against gender-nonconforming individuals is measured. The results show that the X-learner can effectively target treatment regimes and shed light on underlying mechanisms. The X-learner outperforms the T-learner and S-learner in many cases, particularly when the CATE function has a simpler structure or when one treatment group is much larger than the other. The paper also discusses the theoretical properties of the X-learner, including its convergence rates and performance under different assumptions. It shows that the X-learner can achieve faster convergence rates when the CATE function is smoother than the response functions. The X-learner is particularly effective when the number of control units is much larger than the number of treated units, as this allows for more accurate estimation of the CATE function. The paper concludes that the X-learner is a valuable tool for estimating HTE in a wide range of applications, and it provides a software package that implements the methods described.