<aside> 🗺️

Module 5 · Decision Trees - 20 Questions for ML · Supervised Learning · 30-45 min

After this module, you'll understand how decision trees make predictions, why they're interpretable, and how to prevent overfitting.

</aside>

The Setup

You're a data scientist at StreamCart. The support team wants to know why certain customers churn, not just predictions. They need rules they can act on.

Logistic regression gives you a probability, but the PM asks: "Can you give me simple rules like 'If X and Y, then high risk'?"

That's what decision trees do.

The Mental Models

1. The 20 Questions Game

A decision tree is like playing 20 Questions:

• "Is tenure less than 30 days?" → Yes
• "Are support tickets > 3?" → Yes
• "High churn risk!"

Each question splits the data into two groups. After enough splits, you reach a prediction.

Key insight: Trees learn WHICH questions to ask and WHERE to put the thresholds by finding splits that best separate the classes.

2. The Recursive Partition

Trees divide the feature space into rectangular regions:

• First split: One vertical or horizontal line
• Second split: Another line within one region
• Keep splitting until regions are "pure" (mostly one class)