Support for C5.0 decision tree algorithm #30669

Scikit-learn's built-in tree classifier is CART-based; it does not implement C5.0, and selecting criterion="entropy" or "log_loss" does not turn it into C5.0. The algorithms differ in tree construction/pruning and in C5.0 features such as rule sets, winnowing, and its boosting procedure.

I am not aware of an accepted roadmap item to add C5.0 to scikit-learn. If exact C5.0 behavior is required, the established implementation is the C50 package in R; it can be called from a separate R step or through an interoperability layer, with the environment/version pinned for reproducibility.

If the goal is predictive performance rather than reproducing C5.0 specifically, compare DecisionTreeClassifier, RandomForestClassifier, HistGradientBoostingClassifier, and external gradient-boosted tree libraries under the same cross-validation protocol.

Adding C5.0 would be substantially more than another split criterion. It would need a maintained implementation, estimator-API compatibility, sample/class-weight semantics, sparse/missing-data behavior, tests against a trustworthy reference, documentation, and maintainer agreement on long-term scope. A feature proposal would therefore be the right first step before implementation.