c50py — C5.0‑style Decision Trees for Python (clean 0.2.0)

c50py provides transparent, easily inspectable decision trees modelled on Quinlan’s C5.0 algorithm. Both classification and regression trees are supported and expose a scikit‑learn‑like API. The implementation is written from scratch in pure Python/Numpy and includes support for numeric and categorical variables, missing values, pre‑ and post‑pruning, boosting, rule tracing/export and tree drawings identical to scikit-learn's (plot_tree, export_graphviz, export_text).

What's new in 0.4 (0.4.0 and 0.4.1)

Trees are drawn exactly like scikit-learn. c50py now has plot_tree, export_graphviz and export_text with the same parameters, defaults, colours, box text and layout as sklearn.tree.plot_tree, sklearn.tree.export_graphviz and sklearn.tree.export_text:

import matplotlib.pyplot as plt
from c50py import C5Classifier, plot_tree

clf = C5Classifier(categorical_features=["region"]).fit(X, y)

fig, ax = plt.subplots(figsize=(14, 6))
clf.plot_tree(class_names=["stays", "leaves"], filled=True, rounded=True, ax=ax)   # method form
plot_tree(clf, class_names=["stays", "leaves"], filled=True, rounded=True)         # function form, as in sklearn

All of scikit-learn's options work: max_depth, feature_names, class_names, label, filled, impurity, node_ids, proportion, rounded, precision, ax, fontsize (and, for export_graphviz, out_file, leaves_parallel, rotate, special_characters, fontname). Two extras: tree_index, to draw any tree of a boosted model (trials > 1), and palette. Unlike scikit-learn, c50py colours the boxes by default with its own palette (teal, violet, gold, ...), so a C5.0 tree is easy to tell apart from a CART tree; palette="sklearn" gives scikit-learn's orange/blue exactly, and filled=False gives white boxes.

What is specific to C5.0 inside the boxes: the impurity is the entropy (C5.0's criterion; regression trees show squared_error as in scikit-learn), categorical splits read feature in {a, b}, and samples can be fractional when there are missing values (C5.0 sends those cases down both branches).

export_graphviz no longer needs the graphviz package to produce the DOT text, and old calls such as export_graphviz("tree", format="png") keep working.

What's new in 0.3.0

Version 0.3.0 brings the classifier much closer to Quinlan's C4.5/C5.0 and makes it faster:

• Pessimistic pruning as in C4.5. Leaves are penalised with the binomial upper limit (AddErrs from prune.c) and the original confidence-factor table (cf = 0.25 gives a deviate of about 0.69). Pure leaves now receive a positive penalty, so tiny leaves that only memorised the data are pruned. The previous normal approximation gave pure leaves zero penalty, so trees with pure leaves were never pruned, and the old "global" pass could collapse a useful tree into a single leaf on imbalanced data. global_pruning is now off by default.
• Split selection as in C4.5. Gain ratio competes only among features whose gain is at least the average gain (gain_ratio_avg_gain=True); unknown cases count as an extra branch in the split information; and continuous attributes pay the MDL penalty of C4.5 Release 8 (mdl_penalty=True).
• Exhaustive, vectorised numeric thresholds. numeric_threshold_strategy="all" is now the default and is evaluated with cumulative sums, so it is faster than the old 32-quantile subsample.
• pandas DataFrames are accepted directly, column names become feature names, and infer_categorical=True is honoured by the classifier (object, category, bool and string columns). print_tree, export_rules, predict_rule and export_graphviz default to the names seen in fit.
• Boosting. A perfect first tree no longer gets an arbitrary weight of 10; it gets a finite AdaBoost weight and boosting stops, as in C5.0. Combined with the new pruning, trials > 1 now works from min_samples_leaf=1.

Features

• Scikit-learn API: fit(X, y), predict(X), score(X, y).
• Categorical support: Pass categorical_features=[0, 2] to handle categories natively.
• Sample weights: Supports sample_weight in fit for weighted splitting and pruning.
• Missing values: Handles missing values using C5.0's fractional propagation strategy.
• Boosting: Set trials=10 to train a boosted ensemble.
• Rule export: call export_rules() to get a list of human-readable rules.
• Tree drawing like scikit-learn: plot_tree() (matplotlib), export_graphviz() (DOT) and export_text(), with the same parameters and look as sklearn.tree.
• Pretty printing: call print_tree to display the learned splits in a readable nested if/else format (single trees only).

Documentation

For a comprehensive guide on how to use c50py, including advanced features and examples, please see the Usage Guide.

Installation (development mode)

Install the package into your environment in editable mode:

pip install -e .

Quickstart (Classification)

import pandas as pd
from time import perf_counter
from c50py import C5Classifier

df = pd.read_csv("titanic.csv")
t0 = perf_counter(); clf.fit(X, y); print(f"fit: {perf_counter()-t0:.3f}s")

# Inspect the tree
clf.print_tree(feature_names=features, class_names=["No", "Yes"])

# Extract rules for each sample
rules = clf.predict_rule(X, feature_names=features)
print(rules[:5])

# Draw it exactly like sklearn.tree.plot_tree
clf.plot_tree(feature_names=features, class_names=["No", "Yes"], filled=True)

# Export as Graphviz
path = clf.export_graphviz(
    "titanic_tree",
    feature_names=features,
    class_names=["No", "Yes"],
    format="dot"  # save a .dot file directly
)
print(f"DOT file written to {path}")

Quickstart – Regression (Diabetes)

Fit a regression tree to the diabetes dataset and obtain a visualisation:

import pandas as pd
from time import perf_counter
from c50py import C5Regressor

df = pd.read_csv("diabetes.csv")
y = df["target"].values
X_df = df.drop(columns=["target"])
X = X_df.values.astype(object)
features = list(X_df.columns)

reg = C5Regressor(
    min_samples_split=30,
    min_samples_leaf=10,
    pruning=True, cf=0.25, global_pruning=True,
    feature_names=features,
    random_state=42,
    infer_categorical=False, int_as_categorical=False,
    numeric_threshold_strategy="quantile", max_numeric_thresholds=64
)

start = perf_counter(); reg.fit(X, y); print(f"fit: {perf_counter()-start:.3f}s")

# Export to DOT (Graphviz installed optional)
dot_path = reg.export_graphviz("diabetes_tree", feature_names=features, format="dot")
print(f"Tree saved to {dot_path}")

# Export human‑readable rules (single trees only)
rules = reg.export_rules(feature_names=features)
print(rules[:3])

Performance tuning

Several hyperparameters influence model complexity and performance:

• numeric_threshold_strategy ('quantile' | 'all'): subsample candidate numeric thresholds. With 'quantile' the number of splits considered is limited to max_numeric_thresholds per feature per node. 'all' evaluates every unique midpoint (slower on large datasets).
• max_numeric_thresholds: number of candidate thresholds when using 'quantile' (typically 32–64).
• categorical_features: list of names or indices marking categorical columns.
• max_categories_exhaustive: maximum cardinality for exhaustive subset search on categorical features; beyond this a simpler one‑vs‑rest strategy is used.
• infer_categorical/int_as_categorical: enable automatic detection of categorical/boolean/integer columns when dtype information is not explicit.
• max_depth: optional depth limit for extremely noisy or deep trees.

When boosting (trials > 1) the same hyperparameters apply to each base tree.