Regressão com árvores de decisão

A decision tree predicts numbers by asking yes-or-no questions about your features, splitting the data step by step. No straight line required. 🌳

Each branch is a split, like is size above 80? The tree keeps asking, narrowing rows into smaller and smaller groups.

At the bottom sit the leaves. Each leaf predicts one number: the average target of all training rows that landed there.

The tree tries many split points and keeps the one that most reduces prediction error in the resulting groups. Greedy, one step at a time.

The familiar fit-predict contract holds. Create a DecisionTreeRegressor, fit it, and predict just like any other estimator.

Because each leaf gives one flat value, a tree's predictions form a staircase, not a smooth curve. That is fine for many real patterns.

Splits can bend any direction, so a tree captures non-linear relationships that a line would completely miss. That is its real strength.

Left unchecked, a tree grows until each leaf holds one row, memorizing noise. That deep tree overfits and fails on new data.

Control growth with max_depth. A shallower tree stays general and is far easier to read and trust.

You can also set min_samples_leaf so leaves keep enough rows. This stops the tree from carving out tiny, noisy groups.

Trees split on thresholds, so feature units do not matter. You can skip scaling, a nice contrast to Ridge and Lasso.

Think about how a tree turns rows into a prediction.

A decision tree splits data with questions, predicts a leaf average, and bends to non-linear shapes. Cap its depth so it generalizes. 🎯