Supported models and backend selection

This matrix describes the release dependency series: skgrad 0.1.5+, TreeIG 0.2.x, and CBaseline 0.1.2+. All three install with UnifiedIG. Frameworks in the last column are separate optional installations. All models must be fitted or otherwise ready for inference. Inputs must permit numeric interpolation.

Models / interface	Backend and integration	Explained output / limits	Extra installation
LinearRegression, Ridge, Lasso, ElasticNet, LinearSVR	skgrad, exact affine	Regression predictions, including supported multi-output estimators	None
LogisticRegression, RidgeClassifier, LinearSVC	skgrad, exact affine	Binary decision margin; multiclass centered decision scores	None
SVR, NuSVR, binary SVC, NuSVC	skgrad, analytic derivatives and quadrature; linear kernels exact	Built-in linear, polynomial, RBF, sigmoid kernels; no multiclass kernel SVM or precomputed/callable kernels	None
MLPRegressor, MLPClassifier	skgrad and quadrature	Identity regression output; binary logits / centered multiclass logits; identity, logistic, tanh, ReLU hidden activations	None
sklearn Pipeline ending in a skgrad-supported estimator	skgrad chain rule; exact affine/polynomial shortcuts when applicable	StandardScaler, RobustScaler, MaxAbsScaler, MinMaxScaler, PCA, PolynomialFeatures, supported fitted feature selectors, nested pipelines; no arbitrary transformer derivatives	None
DecisionTreeRegressor, RandomForestRegressor, ExtraTreesRegressor, GradientBoostingRegressor	TreeIG, exact split crossings	Scalar regression prediction; numeric splits	None
GradientBoostingClassifier	TreeIG, exact split crossings	Binary margin / centered multiclass margins	None
XGBRegressor, XGBClassifier, native XGBoost Booster	TreeIG, exact split crossings	Supported numeric tree objectives; regression output or raw class margins; TreeIG parser restrictions apply	xgboost
LGBMRegressor, LGBMClassifier, native LightGBM Booster	TreeIG, exact split crossings	Supported numeric tree objectives; regression output or raw class margins; TreeIG parser restrictions apply	lightgbm
HistGradientBoostingRegressor, HistGradientBoostingClassifier; numeric CatBoost models	TreeIG numerical event detector, explicit fallback="tree_numeric"	Prediction / raw class scores; approximate crossing detection	catboost for CatBoost
DecisionTreeClassifier, RandomForestClassifier, ExtraTreesClassifier	TreeIG numerical event detector, explicit fallback="tree_numeric"	Log odds / centered log probabilities of the aggregated model; explicit probability_floor required for zero probabilities	None
torch.nn.Module	Native PyTorch autograd and quadrature	Scalar, raw class scores, or declared multi-output regression; one input tensor per sample	torch
uig.JaxModel(function, ...)	Native JAX autodiff and quadrature	Differentiable batched function; optional explicit parameters	jax
uig.TensorFlowModel(function, ...), TensorFlow-backed Keras	Native TensorFlow gradients and quadrature	Differentiable model output	tensorflow
Keras 3 model	Native configured TensorFlow / JAX / PyTorch backend	Visible sigmoid/softmax classification heads rejected; return logits	keras and its selected backend
Other smooth sklearn regressors / decision-score classifiers	Central finite differences and quadrature, explicit fallback="finite_difference"	Requires predict / decision_function; no probability-only classifiers, multiclass pairwise SVM scores, or tree fallback	Model’s own dependencies

Specialized backends take precedence over an explicitly requested fallback. Unsupported models raise an error; fallback is never silently enabled. TreeIG support depends on objective, split type, and model configuration, not only the Python class. Native categorical inputs and arbitrary discontinuous callables are outside the supported interpolation contract.

For framework vectors, classification is the default: two scores become score[1] - score[0], while three or more are centered across classes. Use output_kind="regression" for multi-output regression. A scalar framework output is explained as supplied; callers must supply a score rather than a probability for classification. UnifiedIG cannot detect every probability transformation hidden inside arbitrary user functions.

A baseline vector is one point; a matrix is a shared distribution, even when its row count equals the input batch size. CBaseline objects supply rows and weights directly. Completeness reconstructs the explained output from the weighted mean baseline output plus summed attributions. Numerical completeness does not guarantee accurate individual allocations; refine quadrature or tree resolution when needed. See semantics for shapes and loss output conventions, and feature spaces for attribute_after.