algorithms/

Every learning algorithm TuiML implements.

Grouped into thirteen families by what the algorithm does, so a wrapped scikit-learn or CapyMOA estimator sits in the same place as its native counterpart. Every class here is registered in the hub under its own name, which is how train and the MCP tools reach it — you name the algorithm, you do not import it.

Supervised learning

trees — decision trees and forests. The strongest

default on tabular data: RandomForestClassifier, DecisionTreeClassifier, DecisionStumpClassifier.

gradient_boosting — boosted ensembles wrapping

XGBoost, LightGBM and CatBoost. Usually the accuracy ceiling, at the cost of tuning.

ensemble — meta-learners that combine other

algorithms: BaggingClassifier, AdaBoostClassifier, StackingClassifier, VotingClassifier.

linear — the interpretable baseline:

LinearRegression, LogisticRegression, SGDClassifier.

svm — maximum-margin classifiers with pluggable

kernels: SVC, SVR.

bayesian — probabilistic models:

NaiveBayesClassifier, NaiveBayesMultinomialClassifier, GaussianProcessesRegressor.

neighbors — instance-based learning, which stores

the training data instead of building a model: KNearestNeighborsClassifier, KNearestNeighborsRegressor.

rules — ZeroRuleClassifier, the majority-class

baseline every other model has to beat.

neural — MultilayerPerceptronClassifier.

Unsupervised learning

clustering — grouping without labels:

KMeansClusterer, DBSCANClusterer, AgglomerativeClusterer, GaussianMixtureClusterer.

anomaly — finding the rare rows:

IsolationForestDetector, LocalOutlierFactorDetector.

associations — co-occurrence rules:

AprioriAssociator, FPGrowthAssociator.

Sequential data

timeseries — forecasting, where order matters and

ordinary cross-validation leaks the future into the past: ARIMA, ExponentialSmoothing, STLDecomposition.

Notes

Names are addressed through the hub, and the namespace is what disambiguates them: "RandomForestClassifier" is the native implementation, "sklearn.RandomForestClassifier" the wrapped one. Both can appear in the same benchmark.

python

>>> import tuiml
>>> model = tuiml.train({
...     "model": {"name": "RandomForestClassifier",
...               "params": {"n_estimators": 50}},
...     "data": {"source": "iris", "target": "class"},
...     "evaluation": {"cv": 5, "metrics": ["accuracy_score"]},
... })
>>> round(model.metrics_["cv_accuracy_score_mean"], 1)
1.0