Skip to content

Get the real datasets

Clone the repository and run the examples from its root. The examples use pandas and matplotlib. Use the repository checkout for examples that import a helper from the examples directory. Install Graphviz if you want to regenerate the optional fitted-tree diagrams.

uv sync
uv run python -c "import pandas, matplotlib, mltreelib"

For an OpenML source below, open its dataset page and download the Parquet file. The parquet_url field in the OpenML data description gives its direct location. Save it under the listed path. These files are excluded from Git.

Dataset Source Local path Study
Diamonds OpenML 46923 datasets/diamonds.parquet Regression
Give Me Some Credit OpenML 46929 datasets/GiveMeSomeCredit.parquet Classification
Wine Quality OpenML 46964 datasets/wine_quality.parquet Clustering
Statlog Shuttle UCI 148 datasets/external/shuttle/ Anomaly detection and archive setup
FLCHAIN Rdatasets CSV datasets/external/real_eval/flchain.csv Survival
Hillstrom CSV mirror datasets/external/real_eval/hillstrom.csv Uplift and source checks
APS Failure OpenML 46908 datasets/APSFailure.parquet Missing sensor data
Concrete strength OpenML 46917 datasets/concrete_compressive_strength.parquet Local models and rules
Miami housing OpenML 46942 datasets/miami_housing.parquet Distributional splits

For the full TabArena collection, see the dataset manifest.

Secondary example: bank churn

Use OpenML 46911 at datasets/Bank_Customer_Churn.parquet to compare three interfaces on the same mixed feature schema. The script fits a churn classifier, a cluster tree and an anomaly tree. Only the classifier uses churn labels.

uv run python examples/real_data_workflow.py

Held-out churn precision–recall and probability diagnostic

This split uses 6,000 training, 2,000 validation and 2,000 test rows. The classifier's test Brier score is 0.1215, compared with 0.1639 for a constant training-rate prediction. The unsupervised outputs describe this customer table. A churn outcome cannot validate an anomaly ranking or determine whether a segment is actionable.