How do Decision Trees handle continuous vs categorical features internally?

Decision Trees handle continuous features via threshold splits and categorical features via encoding or grouping strategies depending on implementation.

For continuous features, decision trees evaluate potential split points by sorting values and testing thresholds like x <= t. For categorical features, implementations differ: CART (used in sklearn) typically requires one-hot encoding, while algorithms like C4.5 can directly split categories or group them optimally. The key challenge is finding a partition of categories that maximizes impurity reduction, which becomes combinatorially expensive for high-cardinality features. This is why modern systems often encode categorical variables or use specialized tree variants.