imodels.viz: nice viz for interpretable models

Chandan Singh · October 2026

Visualization is critical to understanding fitted models. imodels.viz lets you visualize any imodels model and most sklearn models nicely. It produces either a static figure (SVG, PNG, PDF) or an interactive page (for html or notebooks).

Quickstart

from imodels import FIGSClassifier
from imodels import viz
from sklearn.datasets import load_breast_cancer

X, y = load_breast_cancer(return_X_y=True, as_frame=True)
model = FIGSClassifier(max_rules=8).fit(X, y)

viz.draw(model, X, y).save("figs.svg")             # static: .svg .png .pdf .html
viz.interactive(model, X, y).save("figs.html")     # one offline page; also renders inline in Jupyter
print(model)                                       # the model as readable text

Both calls take the fitted model and, optionally, the training data. With data, every split shows the distribution of its feature and every leaf its class mix or target range; without it, the figure falls back to what the model stores. Other arguments set names (feature_names, class_names, target_name), highlight one sample's path (x=), limit the depth drawn (max_depth), and switch theme, orientation and style. See the API docs.

1. Supported models

Decision trees10

DecisionTreeClassifier·DecisionTreeRegressor·ExtraTreeClassifier·ExtraTreeRegressor·GreedyTree·DecisionTreeCCP·HSTree·TaoTree·C45Tree·FastSmallTree

Tree ensembles6

RandomForest·ExtraTrees·GradientBoosting·HistGradientBoosting·FIGS·IRF

Rule lists4

GreedyRuleList·OneR·FastFrugalTree·BayesianRuleList

Rule sets7

RuleFit·FPLasso·SkopeRules·FPSkope·BoostedRules·Slipper·BayesianRuleSet

Scoring systems2

FastRiskScore·SLIM

Additive and linear models12

LinearRegression·Ridge/Lasso/ElasticNet·LogisticRegression·SGD·LinearSVC/LinearSVR·Huber/RANSAC·Poisson/Gamma/Tweedie·RidgeClassifier/Perceptron·IsotonicRegression·TreeGAM·GPGam·MarginalShrinkageLinear

scikit-learnimodels · Scaling pipelines are drawn in raw units. Not supported: models without readable structure (kNN, kernel SVMs, MLPs).

Every tree is a scikit-learn tree, so dtreeviz works too

imodels.viz has no drawing code for any particular imodels tree. Every tree-based model is first exported to the scikit-learn estimator that makes the same predictions, with imodels.to_sklearn, and then drawn like any scikit-learn model: single trees (CART variants, HSTree, TAO, C4.5, FastSmallTree) become a DecisionTreeClassifier or DecisionTreeRegressor, IRF becomes a RandomForestClassifier, and FIGS becomes a list of regression trees whose predictions add up, the same view as gradient boosting. The tests check every export against the model's own predict / predict_proba.

import imodels
tree = imodels.to_sklearn(model, X)    # X (optional) recounts each node's samples

# anything that reads scikit-learn trees now reads imodels trees, e.g. dtreeviz
import dtreeviz
dtreeviz.model(tree, X, y, feature_names=list(X.columns)).view()
sklearn.tree.plot_tree(tree)

So the export also makes every tree-based imodels model work with dtreeviz, including ones dtreeviz could not read before, such as C4.5, FastSmallTree and IRF.

2. Interactive mode

viz.interactive writes one self-contained HTML file with no server and no network requests. Across all models the page offers the same tools:

Try a sample. Type feature values or load a random training row. The path it takes lights up, and a waterfall shows how the prediction is built (leaf values, rule weights, points or shape-function terms).
What would change it. The smallest single-feature and two-feature changes that flip a class, or move a regression prediction the most, with round values just past each threshold.
Features panel. Each feature's importance and marginal distribution; click one to highlight every split or rule that uses it.
Fold and simplify. Click a split to fold its subtree; the Simple toggle swaps the charts for boxes filled by class proportions. Export the current view as SVG.