This article introduces the X-learner, a new meta-algorithm for estimating heterogeneous treatment effects (CATE) using machine learning. The X-learner builds on base algorithms like Random Forests (RF) and Bayesian Additive Regression Trees (BART) to estimate the CATE function, which represents the average treatment effect conditional on covariates. The X-learner is designed to be efficient when one treatment group is much larger than the other and can exploit structural properties of the CATE function. It performs well in simulations and is particularly effective when the CATE is smooth or linear. The article compares the X-learner with other meta-learners like the T-learner and S-learner, showing that the X-learner often outperforms them, especially in unbalanced designs. The X-learner is also shown to be more efficient in cases where the CATE has a simpler structure. The paper includes two real-world applications: one on voter turnout and another on reducing transphobia through door-to-door canvassing. The X-learner is implemented in a software package and is recommended for use when the CATE is expected to be smooth or when one treatment group is much larger than the other. The article also provides theoretical results showing that the X-learner can achieve faster convergence rates under certain conditions, making it a valuable tool for estimating heterogeneous treatment effects in experimental and observational studies.This article introduces the X-learner, a new meta-algorithm for estimating heterogeneous treatment effects (CATE) using machine learning. The X-learner builds on base algorithms like Random Forests (RF) and Bayesian Additive Regression Trees (BART) to estimate the CATE function, which represents the average treatment effect conditional on covariates. The X-learner is designed to be efficient when one treatment group is much larger than the other and can exploit structural properties of the CATE function. It performs well in simulations and is particularly effective when the CATE is smooth or linear. The article compares the X-learner with other meta-learners like the T-learner and S-learner, showing that the X-learner often outperforms them, especially in unbalanced designs. The X-learner is also shown to be more efficient in cases where the CATE has a simpler structure. The paper includes two real-world applications: one on voter turnout and another on reducing transphobia through door-to-door canvassing. The X-learner is implemented in a software package and is recommended for use when the CATE is expected to be smooth or when one treatment group is much larger than the other. The article also provides theoretical results showing that the X-learner can achieve faster convergence rates under certain conditions, making it a valuable tool for estimating heterogeneous treatment effects in experimental and observational studies.