A machine learning model that makes predictions by asking a sequence of yes-or-no questions about input features, splitting data at each step based on which question best separates the outcome classes. Decision trees are among the most interpretable models available, which makes them valuable wherever a client needs to understand why a prediction was made.
Also known as classification tree, regression tree, decision tree model
A decision tree is a model shaped like a flowchart. At each internal node, the tree asks a question about an input feature: “Is the customer’s account age greater than 90 days?” Depending on the answer, the tree branches left or right. After enough branching steps, the tree reaches a leaf node that produces a prediction: qualified, high-risk, likely-to-churn.
Trees are trained by finding the sequence of questions that best separates the training examples by their target category at each split. The algorithm evaluates all possible features and all possible split points and selects the combination that produces the purest separation. This process repeats recursively until the tree reaches a stopping criterion, such as a maximum depth or a minimum number of examples per leaf.
Single decision trees tend to overfit on training data if allowed to grow too deep. The most common practical fix is to build many trees on different random subsamples and average their predictions, an approach called a random forest, or to build trees sequentially where each new tree corrects the errors of the previous ones, an approach called gradient boosting. Both methods trade the interpretability of a single tree for substantially better predictive performance.
Agency work involves presenting analytical recommendations to clients who are not data scientists. Decision trees and their close relatives are among the few model types that produce outputs a non-technical audience can actually follow. Showing a client the sequence of rules the model uses to classify a lead as high-value is a fundamentally different conversation from showing them a black-box score.
Interpretability is sometimes a business requirement. Some clients in regulated industries need to be able to explain why an AI system made a specific decision about a specific individual. Decision tree-based models can satisfy that requirement in a way that deep learning models cannot. Knowing when interpretability requirements apply changes the model selection conversation.
They are useful for discovering actionable segmentation rules. A decision tree trained on customer value data can surface a segmentation rule in plain language: “Customers who made a purchase in the first 14 days and had more than three site sessions before converting are high-value at a rate four times higher than the baseline.” That rule is immediately actionable for a campaign team without requiring them to understand how the model works.
Ensemble methods built on trees power many practical ML tools. Random forests and gradient boosted trees, XGBoost being the most widely used, are among the best-performing models on tabular data. Many scoring tools, attribution models, and predictive analytics platforms use these methods under the hood. Understanding the single-tree foundation makes the ensemble methods easier to reason about.
An agency builds a lead qualification model for a B2B software client and uses a gradient boosted tree model for its performance, but also trains a single, shallow decision tree for interpretation. The shallow tree reveals that the top splitting rule is company size (fewer than 200 employees branches toward qualified; above does not), followed by industry vertical, followed by whether the lead came from an organic search or a paid channel. The agency presents this segmentation logic to the client’s sales team, who recognize it immediately as consistent with their manual qualification criteria and gain confidence in the more complex model’s recommendations.
The generative AI foundations module of the workshop covers how today’s models work, what they can and can’t do, and how to choose between them for the specific agency and client use cases where interpretability matters.