Tree-based methods are a workhorse for e-commerce risk scoring because they perform strongly on structured data and remain practical to operate at scale. This chapter begins with logistic regression as an interpretable baseline and then develops decision trees as flexible learners that capture nonlinear interactions common in fraud and abuse. It proceeds to ensembles-bagging and, especially, gradient boosting-which have become production standards for tabular risk modeling. Practical topics include regularization, early stopping, handling missing values, monotonic constraints, and feature interactions under noisy signals. Because risk decisions are frequently appealed, audited, or operationally contested, the chapter also addresses interpretability and operator trust, including feature importance and explanation strategies that remain robust in production. Throughout, model choice is tied to engineering constraints such as latency, stability, and ease of iteration. We highlight common modeling pitfalls such as single-feature brittleness and stale retraining and show how ensembles support robust decisions when data is noisy, sparse, or strategically manipulated. Concrete guidance is provided on where trees fit best, when to prefer rules, and how to combine both in a single decision pipeline. The chapter positions tree-based models as a foundational component of scalable decision engines that integrate cleanly with rules, human review, and policy controls.