IMoTE: Interactive Model Tree Explorer

IMoTE is an interactive application to fit, visualise and explore linear model trees for regression on tabular data. It shows a fitted tree as an interactive graph, lets you look at the linear model and the data behind every node, and traces a single data point through the tree to explain its prediction.

Why IMoTE?

A linear model tree combines the rule based structure of a decision tree with linear models in its leaves. That usually makes it more accurate than a classic regression tree of the same depth. In practice, these trees are often hard to read from code or a text dump. IMoTE makes them easy to explore, explain and present.

Features

Fit or load a tree

• Fit PILOT [1] or M5 [2] trees on built-in PMLB datasets or on your own CSV file.
• Save the tree you are looking as an SVG image or to load it again later.
• Load trees fitted with other tools through an adapter. An adapter for trees from the R package partykit is included as an example.

Edit and explore the tree

• Pan, zoom and click through the tree as an interactive graph.
• Collapse and expand branches, collapse the whole tree to a given level, show a single branch as a subtree, or prune the tree (PILOT only).
• Switch between normal, minimal and tiny nodes or color nodes by the feature they use.
• Show the data flow: edge width or node size scale with the number of training samples that pass through them.

Inspect nodes and predictions

• Click a node to see its type, split and metrics (number of samples, RSS, MAE, ...).
• See the plot of any node: the data and the fit of a split, or the prediction plot of a leaf that shows how each feature contributes to its predictions.
• Turn on Show all node plots to replace every node of the graph by its plot.
• Highlight the path of a data point through the tree, with its prediction.

Installation

IMoTE needs Python 3.10 or newer.

pip install imote

Start the application with:

imote

Then open the address shown in the terminal (by default http://127.0.0.1:8050). The app starts with an example PILOT tree. Click a node to inspect it, or open the New Tree card on the right to fit a tree on another dataset.

IMoTE stores saved trees, uploaded datasets and generated plots in ~/.imote. Set the IMOTE_HOME environment variable to use another folder.

Documentation

The full documentation is at https://stan-uantwerp.github.io/IMoTE/, it includes:

• the User Guide, with a page for every part of the interface,
• the Gallery, with screenshots of what IMoTE can show,
• a guide on how to write your own adapter for another type of linear model tree,
• an API reference of the tree and node classes used for visualisation.

Development

git clone https://github.com/STAN-UAntwerp/IMoTE
cd IMoTE
pip install -e ".[dev]"   # IMoTE in editable mode, plus pytest, build and twine
python -m imote.app       # start the app in Dash debug mode, it reloads on code changes

To work on the documentation: pip install -e ".[docs]", then mkdocs serve.

Questions and ideas are welcome in Discussions, bugs can be reported as an issue.

References

_{[1] Raymaekers, J., Rousseeuw, P. J., Verdonck, T., & Yao, R. (2024). Fast linear model trees by PILOT. Machine Learning, 113(9), 6561-6610. https://doi.org/10.1007/s10994-024-06590-3 (implementation: fast-model-trees)}

_{[2] Quinlan, J. R. (1992). Learning with continuous classes. In 5th Australian Joint Conference on Artificial Intelligence (Vol. 92, pp. 343-348). (implementation: m5py)}