Skip to content

Choosing a model

highFIS exposes several TSK families behind the same scikit-learn interface, so you can swap one for another by changing a single class. A few rules of thumb:

  • TSKClassifier — the vanilla TSK baseline (product T-norm, sum normalization).
  • HTSKClassifier — high-dimensional TSK (geometric-mean / log-space) for more features.
  • LogTSKClassifier — log-domain inverse-log normalization for stable aggregation.
  • ADPTSKClassifier — adaptive double-parameter softmin, tuned for high-dimensional data.
  • DGTSKClassifier / FSREADATSKClassifier — add embedded feature selection and rule extraction (see the feature-selection recipe).

Because the estimators are scikit-learn compatible, you can let GridSearchCV select the model for you by putting the estimator itself in the search space — searching over both the model family and its hyperparameters in one pass:

from sklearn.datasets import load_iris
from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import MinMaxScaler

from highfis import HTSKClassifier, LogTSKClassifier, TSKClassifier

X, y = load_iris(return_X_y=True)

pipe = Pipeline([("scale", MinMaxScaler()), ("clf", TSKClassifier())])

# Each grid entry pins the "clf" step to a model family and searches its n_mfs.
param_grid = [
    {"clf": [TSKClassifier(mf_init="grid", epochs=40, random_state=0)], "clf__n_mfs": [2, 3]},
    {"clf": [HTSKClassifier(mf_init="grid", epochs=40, random_state=0)], "clf__n_mfs": [2, 3]},
    {"clf": [LogTSKClassifier(mf_init="grid", epochs=40, random_state=0)], "clf__n_mfs": [2, 3]},
]

search = GridSearchCV(pipe, param_grid, cv=3)
search.fit(X, y)

print("best model:", type(search.best_params_["clf"]).__name__)
print("best n_mfs:", search.best_params_["clf__n_mfs"])
print("best cv score:", round(float(search.best_score_), 3))

best model: LogTSKClassifier
best n_mfs: 3
best cv score: 0.967

GridSearchCV refits the winning configuration on the full data as search.best_estimator_, ready to predict. For larger or continuous search spaces, use RandomizedSearchCV the same way. If you only want a quick side-by-side comparison without selecting/refitting, a cross_val_score loop over the estimators also works.

The adaptive/gated families (ADPTSKClassifier, DGTSKClassifier, FSREADATSKClassifier) are designed for high-dimensional problems and shine there rather than on a small dataset like Iris. Every family also has a *Regressor counterpart with the same interface.