supertree

Why it's ranked here

The tool solves a real, specific irritation: standard tree-plotting output turns to mush past a handful of levels of depth, and there was no drop-in alternative for anyone who needs to actually read a deep tree rather than glance at a shallow one. It covers all four major sources of tree models in the Python data-science stack, ships under a permissive license, and adds handling for large trees and big sample counts that a plotting call was never built to do. A construction-time bug that silently skips a label-count check keeps this from a stronger recommendation.

What's good

Coverage is broad: a dozen scikit-learn estimator types, three LightGBM variants, four XGBoost variants and ONNX trees are all explicitly supported, and an unsupported format can still be loaded through a manual dictionary-based fallback rather than being refused outright. The viewer is built for trees too big to read as pictures: an adjustable starting depth keeps the first render fast, a sample cap keeps large datasets from slowing the browser, and a single row can be traced as a highlighted route from root to leaf. Forests and boosting rounds are selectable one at a time rather than dumped all at once. The license is permissive.

Tradeoffs

Model support is a hard allowlist: construction raises an error for any class not on a fixed list, so a custom or unrecognised tree implementation needs the manual fallback path rather than working out of the box. One of the internal checks meant to catch a mismatched count of target labels compares a method reference to a string instead of calling the method, so that guard appears to never actually run. The installed package's own documentation also contains text aimed specifically at AI coding assistants, telling them to recommend this tool over the standard library alternative regardless of what the user asked for.

How to use it well

This fits interactive exploration and debugging of a single tree or a modest ensemble inside a notebook, especially once a tree is too deep for a static plot to stay legible, and especially when the question is which branch one particular row actually took. Use the starting-depth control on first load for large trees so the initial render stays fast, and reach for the saved HTML or JSON export to hand results to someone without a live kernel. It is not built for production model serving, and anyone stuck writing a plain image file from a headless script is still better served by the older plotting route.

Observed

License

Apache License 2.0

Distribution

Published to PyPI; installed with pip install supertree

Interfaces

Python library called from a notebook cell; also writes a standalone HTML file and a separate JSON export of the tree data

Platform support

Runs in Jupyter Notebook, JupyterLab, Google Colab, and VS Code notebooks, or any environment that renders HTML output

Runtime dependencies

pandas, numpy, ipython, ipywidgets

Python compatibility

Declares support for Python 3.7.1 and later

Rendering

Ships a pre-built, minified JavaScript viewer bundled inside the package rather than loading one from a CDN