Categorical Features & Target Encoding

shinrin handles categorical columns via CatBoost-style ordered-less target encoding: categories are replaced by smoothed per-category target statistics, and trees train on ordinary numeric thresholds. Because encoded-threshold splits always separate prefixes of the encoding-sorted categories, the original categorical split can be recovered exactly afterwards — for interpretability (CategoricalTree) and for deployment (BRANCH_MEMBER ONNX export).

TargetEncoder

import numpy as np
import shinrin

# column 0 holds integer category codes, column 1 is numeric
enc = shinrin.TargetEncoder(categorical_features=[0], smoothing=1.0)
X_enc = enc.fit_transform(X_raw, y)

model = shinrin.MondrianForestRegressor(n_estimators=20).fit(X_enc, y)

Each category c of feature f is encoded as

enc(c) = (sum_y(c) + smoothing * prior) / (count(c) + smoothing)

where prior is the global target mean. Rare categories shrink toward the prior; unseen categories at transform time map to the prior.

Attributes after fitting:

Attribute	Meaning
categorical_features_	Indices of encoded columns
categories_	Per-column sorted category values
encodings_	Per-column encoded values aligned with categories_
prior_	Global target mean used for smoothing

Recovering categorical splits

Splits on an encoded column partition the categories by their encoded values. Since only prefixes of the encoding-sorted categories are representable as a single threshold, every trained split corresponds to exactly one membership set. to_categorical_tree() recovers it:

ctree = shinrin.to_categorical_tree(model, enc)

# render human-readable rules over RAW inputs
print(ctree.to_text(feature_names=["color", "size"]))
#   x0 in {0.0, 2.0}
#   ├─ x1 <= 1.32  →  10.2
#   └─ ...

# apply() consumes raw pre-encoding samples
leaf = ctree.apply(X_raw)

For forests, to_categorical_tree(model, enc) returns a list (one CategoricalTree per estimator). Each tree also round-trips back to the encoded representation with ctree.to_encoded_thresholds(enc), so both views stay interchangeable.

Exporting to ONNX without the encoder

The recovered partitions let the ONNX export use ai.onnx.ml opset-5 TreeEnsemble BRANCH_MEMBER splits over raw category codes, removing the encoder from the deployed graph entirely. See Categorical features & BRANCH_MEMBER.