FuseKernel

class modeva.models.fusekernel.api.MoFuseKernelRegressor(name=None, use_xgb=True, use_rbf=False, use_spectral=True, tree_depths=None, spectral_params=None, fit_method='grid', solver='nystrom', backend='xgboost', gbdt_params=None, residual_nw=False, max_interaction_features=6, max_interactions=10, interpret_n_grid=30, interpret_ref_size=1000, random_state=0)

fuseKernel fused kernel-ridge regressor for MoDeVa.

Parameters:

namestr, optional

use_xgb, use_rbf, use_spectralbool

Which kernel channels to fuse (defaults: tree + spectral). use_xgb turns on the tree co-membership channel regardless of backend.

tree_depthstuple of int, optional

One co-membership kernel per depth (multi-depth tree fusion).

spectral_paramsdict, optional

MS-SKM kwargs (H / K / kernel / …). None -> the fuseKernel defaults (kernel="laplace", H=4, K=8, solver="nystrom").

fit_method{“grid”, “adam”, “nlml”, “oof”, “gcv”, “sure”}, default=”grid”

Fusion-weight selection (grid/adam are leakage-free, query-scored).

solver{“nystrom”, “auto”, “lanczos”, “matfree”}, default=”nystrom”

How the spectral channel decodes its kernel (only used when use_spectral=True). "nystrom" is the linear-in-n low-rank decode and the fastest/most scalable default; "lanczos" is the exact dense decode (best for small data); "auto" uses dense below ~20k rows and switches to Nystrom above; "matfree" is a matrix-free CG solve. An explicit "solver" key in spectral_params overrides this.

backend{“xgboost”, “lightgbm”, “catboost”}, default=”xgboost”

Gradient-boosted ensemble that defines the leaf co-membership partition.

gbdt_paramsdict, optional

Native params for the chosen backend. None -> that backend’s defaults.

residual_nwpassthrough fuseKernel option.

max_interaction_features, max_interactionsinteraction screening for interpret.

interpret_n_grid, interpret_ref_sizefANOVA grid / reference-sample size.

random_stateint

calibrate_interval(X, y, alpha=0.1, max_depth=5)

fuseKernel’s intervals are the GP posterior – no fitting needed; just record alpha.

channel_contributions(X)

Exact additive per-channel decomposition of the fused prediction (regression).

Returns a ValidationResult: value holds the per-channel contribution arrays and the intercept; table is the mean absolute contribution per channel; plot() shows the bar.

diagnose_weak_clusters(dataset, n_clusters: int = 5, **kw)

Per-cluster train/test performance breakdown of the fitted fused kernel.

Nyström spectral clustering of the model’s own kernel partitions the data into n_clusters regions; the model metric is reported per region on train and test, so regions where the model underperforms (large train/test gap or low headline metric) are exposed. Works for both the classic two-channel and the general (spectral / multi-depth) paths.

Returns a ValidationResult: table is the per-cluster metric breakdown (with an ALL aggregate row); value holds the cluster labels, spectral embeddings and the weakest-cluster ranking; plot() shows the per-cluster test metric as a bar.

fit(X, y, sample_weight=None, feature_names=None)

interpret(dataset)

Build the inherent FANOVA interpreter. Precomputes fuseKernel’s main-effect curves and pairwise-interaction surfaces over a reference sample, then exposes them through the standard InterpretFANOVA so ts.interpret_* work. Needs a spectral channel.

predict(X)

Model predictions, calling the child class’s ‘_predict’ method.

Parameters:

Xnp.ndarray of shape (n_samples, n_features)

Feature matrix for prediction.

Returns:

np.ndarray: The (calibrated) final prediction

predict_dist(X)

GP predictive (mean, variance) in target units.

predict_effect(fidx, X)

Raw prediction of one main effect (len-1 fidx) or pairwise interaction (len-2 fidx).

predict_interaction(X)

Pairwise-interaction raw predictions, shape (n, n_interactions).

predict_interval(X)

Closed-form GP posterior prediction interval at level 1 - alpha.

predict_main_effect(X)

Per-feature main-effect raw predictions, shape (n, n_features).

class modeva.models.fusekernel.api.MoFuseKernelClassifier(name=None, use_xgb=True, use_rbf=False, use_spectral=True, tree_depths=None, spectral_params=None, fit_method='grid', solver='nystrom', backend='xgboost', gbdt_params=None, max_interaction_features=6, max_interactions=10, interpret_n_grid=30, interpret_ref_size=1000, random_state=0)

fuseKernel fused kernel-ridge classifier for MoDeVa (binary and multiclass).

The fused KRR decodes one-hot targets to per-class scores; predict_proba is a temperature-calibrated softmax. Same parameters as MoFuseKernelRegressor.

channel_contributions(X)

Exact additive per-channel decomposition of the fused prediction (regression).

Returns a ValidationResult: value holds the per-channel contribution arrays and the intercept; table is the mean absolute contribution per channel; plot() shows the bar.

decision_function(X)

Computes the decision function for the given input data.

Parameters:

Xnp.ndarray of shape (n_samples, n_features)

Feature matrix for prediction.

calibrationbool, default=True

If True, will use calibrated probability if calibration is done. Otherwise, will use raw probability.

Returns:

logit_predictionarray, shape (n_samples,) or (n_samples, n_classes)

Array of (calibrated) logit predictions.

diagnose_weak_clusters(dataset, n_clusters: int = 5, **kw)

Per-cluster train/test performance breakdown of the fitted fused kernel.

Nyström spectral clustering of the model’s own kernel partitions the data into n_clusters regions; the model metric is reported per region on train and test, so regions where the model underperforms (large train/test gap or low headline metric) are exposed. Works for both the classic two-channel and the general (spectral / multi-depth) paths.

Returns a ValidationResult: table is the per-cluster metric breakdown (with an ALL aggregate row); value holds the cluster labels, spectral embeddings and the weakest-cluster ranking; plot() shows the per-cluster test metric as a bar.

fit(X, y, sample_weight=None, feature_names=None)

interpret(dataset)

Build the inherent FANOVA interpreter. Precomputes fuseKernel’s main-effect curves and pairwise-interaction surfaces over a reference sample, then exposes them through the standard InterpretFANOVA so ts.interpret_* work. Needs a spectral channel.

predict(X)

Model predictions, calling the child class’s ‘_predict’ method.

Parameters:

Xnp.ndarray of shape (n_samples, n_features)

Feature matrix for prediction.

calibrationbool, default=True

If True, will use calibrated probability if calibration is done. Otherwise, will use raw probability.

Returns:

np.ndarray: The (calibrated) final prediction

predict_effect(fidx, X)

Raw prediction of one main effect (len-1 fidx) or pairwise interaction (len-2 fidx).

predict_interaction(X)

Pairwise-interaction raw predictions, shape (n, n_interactions).

predict_main_effect(X)

Per-feature main-effect raw predictions, shape (n, n_features).

predict_proba(X)

Predict (calibrated) probabilities for X.

Parameters:

Xnp.ndarray of shape (n_samples, n_features)

Feature matrix for prediction.

calibrationbool, default=True

If True, will return calibrated probability if calibration is done. Otherwise, will return raw probability.

Returns:

np.ndarray: The (calibrated) predicted probabilities