Estimator Interfaces

Interface	Methods	Description
Estimator	type(): EstimatorType	Root interface for all estimators
Learner	train(Dataset): void · predict(Dataset): Tensor	Supervised and unsupervised learners
Probabilistic	predictProba(Dataset): Tensor	Returns class probability matrix [N, C]
Scoring	score(Dataset): float	Returns accuracy (classifiers) or R² (regressors)
Online	partialFit(Dataset): void	Incremental / streaming training
RanksFeatures	featureImportances(): Tensor	Returns per-feature importance scores
Verbose	setLogger(LoggerInterface): void	PSR-3 training progress logging
Persistable	save(string): void · load(string): static	Checkpoint save/load via SafeTensors

Classifiers

Namespace: Pml\Estimators\Classifiers\

Class	Constructor Parameters	Interfaces	Notes
GBDTClassifier	nEstimators=100, maxDepth=4, numBins=255, lr=0.1, lambda=1.0, minChildWeight=1, subSample=1.0, colSampleByTree=1.0	Scoring RanksFeatures	Histogram-based leaf-wise GBDT (LightGBM-style). C-level trees.
RandomForestClassifier	nEstimators=100, maxDepth=10, minSamplesSplit=2	Scoring	Bagged CART trees, random feature subsets. Parallelized over trees.
ExtraTreesClassifier	nEstimators=100, maxDepth=10, minSamplesSplit=2	Scoring	Extremely randomized trees — random split thresholds, lower variance.
DecisionTreeClassifier	maxDepth=10, minSamplesSplit=2, maxFeatures=null	Scoring	CART with Gini impurity. C-level split search.
LogisticRegression	epochs=100, learningRate=0.01, batchSize=32	Scoring Probabilistic	Binary or multi-class via softmax. Mini-batch SGD.
SoftmaxClassifier	epochs, learningRate, batchSize, lambda	Probabilistic	Multi-class linear classifier with L2 regularization.
SVC	c=1.0, epochs=100, learningRate=0.01, batchSize=32	Scoring	Linear SVM with hinge loss. SGD optimization.
MLPClassifier	hidden=[100], epochs=100, batchSize=32, learningRate=0.001, dropout=0.0	Scoring Probabilistic	Multi-layer perceptron using Sequential internally. ReLU activations.
AdaBoostClassifier	nEstimators=50, learningRate=1.0	Scoring	SAMME algorithm with DecisionTree weak learners.
LogitBoost	nEstimators, learningRate, maxDepth	Probabilistic	Gradient boosting with logistic loss. Probabilistic outputs.
GaussianNB	—	Probabilistic	Gaussian Naive Bayes. C-level log-likelihood computation.
BernoulliNB	alpha=1.0	Probabilistic	For binary feature vectors. Laplace smoothing.
MultinomialNB	alpha=1.0	Probabilistic	For count features (e.g., TF-IDF). Laplace smoothing.
CategoricalNB	alpha=1.0	Probabilistic	For categorical integer-encoded features.
KNNClassifier	k=5	Scoring Probabilistic	Brute-force KNN via C pairwise-L2. O(N·D) per query.
KDNeighborsClassifier	k=5, leafSize=30	Scoring	KD-Tree based KNN. Faster for low-D (<20 features).
RadiusNeighborsClassifier	radius=1.0	Scoring	Classifies using all neighbors within a fixed radius.
OneVsRest	prototype: Learner	Scoring	Wraps any binary classifier for multi-class OVR decomposition.
VotingClassifier	estimators: Learner[]	Scoring	Majority-vote ensemble. Optionally soft-vote with Probabilistic members.
DummyClassifier	—		Returns the most frequent class label. Baseline.

Regressors

Namespace: Pml\Estimators\Regression\

Class	Constructor Parameters	Notes
GBDTRegressor	nEstimators=100, maxDepth=4, numBins=255, lr=0.1, lambda=1.0	Histogram leaf-wise GBDT with MSE loss. Same C kernel as GBDTClassifier.
GradientBoostingRegressor	nEstimators, learningRate, maxDepth, loss	Stage-wise gradient boosting with pluggable loss (MSE, Huber, MAE).
RandomForestRegressor	— (see classifier)	Alias for ExtraTreeRegressor ensemble.
ExtraTreeRegressor	nEstimators, maxDepth, minSamplesSplit	Extremely randomized regression trees.
DecisionTreeRegressor	maxDepth=3, minSamplesSplit=2, maxFeatures=null	CART with MSE criterion.
LinearRegression	—	Closed-form OLS via LAPACKE SGELSD. O(D³) — for small D only.
Ridge	alpha=1.0, epochs=100, learningRate=0.01, batchSize=32	L2-regularized linear regression. Mini-batch SGD.
Lasso	alpha=1.0, epochs=100, learningRate=0.01, batchSize=32	L1-regularized via subgradient SGD. Produces sparse weights.
ElasticNet	alpha=1.0, l1Ratio=0.5, epochs=100, learningRate=0.01, batchSize=32	Combines L1 + L2 regularization.
Adaline	epochs, learningRate, batchSize	Adaptive linear neuron. Mini-batch gradient descent, MSE loss.
MLPRegressor	hidden=[100], epochs=100, batchSize=32, learningRate=0.001	Deep MLP for regression. Outputs continuous values via linear head.
SVR	c, epsilon, kernel, epochs, learningRate	ε-insensitive SVR. Hinge loss SGD with RBF or linear kernel.
KNNRegressor	k=5	KNN regression — average of k nearest target values.
KDNeighborsRegressor	k=5	KD-Tree KNN regression.
VotingRegressor	estimators: Learner[]	Average predictions from an ensemble of regressors.
DummyRegressor	—	Returns training label mean. Baseline.

Anomaly Detectors

Namespace: Pml\Estimators\AnomalyDetectors\

Anomaly detectors implement Learner. predict() returns a Tensor of labels: 1.0 = inlier, -1.0 = outlier.

Class	Constructor Parameters	Notes
IsolationForest	nEstimators=100, sampleSize=256, contamination=0.1	Builds random isolation trees. Anomaly score via average path length. C-accelerated.
LocalOutlierFactor	k=20	LOF score based on local reachability density vs k neighbors.
GaussianMLE	threshold=-10.0	Fits a multivariate Gaussian. Points below log-likelihood threshold are anomalies.
OneClassSVM	nu=0.1, kernel=RBF(0.1), epochs=100, learningRate=0.01	One-class SVM with RBF or linear kernel. SGD optimization.
Loda	nProjections=100, bins=10	Lightweight Online Detector of Anomalies. Random projections + histogram density.
RobustZScore	threshold=3.5	Modified Z-Score using median absolute deviation. Robust to outliers in the reference.

Clusterers

Namespace: Pml\Estimators\Clusterers\

Clusterers implement Learner. predict() returns cluster index labels.

Class	Constructor Parameters	Notes
KMeans	k=3, maxIter=300, tolerance=1e-4	Lloyd's algorithm. Centroid init via KMeans++ seeder. C-level assignment and update.
GaussianMixture	k=3, maxIter=100, tolerance=1e-4	EM algorithm for GMM. Soft cluster assignments. Uses LAPACKE for covariance inversion.
FuzzyCMeans	k=3, fuzziness=2.0, maxIter=300, tolerance=1e-4	Fuzzy / soft clustering. Fuzziness m > 1; higher m = softer boundaries.
DBSCAN	epsilon=0.5, minSamples=5	Density-based. No pre-specified k. Marks noise points as label -1.
MeanShift	bandwidth=1.0, maxIter=100, tolerance=1e-4	Non-parametric; finds cluster centers as density peaks. Automatically determines k.

KMeans Seeders

Pass via seeder: parameter to KMeans or GaussianMixture:

Seeder	Description
PlusPlus	KMeans++ — default, reduces poor initializations
KMC2	Fast KMeans++ approximation — O(k) vs O(Nk)
Random	Uniform random seeding — fast but prone to bad local minima
Preset	Fixed user-supplied centroids

Decomposition & Manifold

Class	Namespace	Parameters	Notes
PrincipalComponentAnalysis	Estimators\Decomposition\	nComponents=2	Full SVD via LAPACKE. Returns loadings, explained variance ratio.
TSNE	Estimators\Manifold\	nComponents=2, perplexity=30, lr=200, maxIter=1000	t-SNE via Barnes-Hut approximation. For visualization. O(N log N) per iter.

Meta Estimators

Class	Namespace	Description
GridSearch	Estimators\Meta\	Exhaustive hyperparameter search over a parameter grid. Trains one model per combination.
RandomSearch	Estimators\Meta\	Random hyperparameter sampling — faster than grid for large spaces.
PlattScaler	Estimators\Meta\	Wraps any classifier and calibrates probability outputs via Platt scaling (logistic regression on decision scores).
BootstrapAggregator	Pml\	Bagging meta-estimator. Wraps any Learner, trains N copies on bootstrap samples, averages/votes predictions.
CommitteeMachine	Pml\	Weighted ensemble of experts. Weights are updated based on per-expert validation error.
StackingRegressor	Pml\Ensemble\	Stacked generalization: base estimators → meta-regressor trained on OOF predictions.

Usage Example

use Pml\Estimators\Classifiers\GBDTClassifier;
use Pml\Dataset;

// Train
$clf = new GBDTClassifier(
    nEstimators: 300,
    maxDepth:    6,
    lr:          0.05,
    lambda:      1.0,
    subSample:   0.8,
);
$clf->train($trainSet);

// Predict
$preds = $clf->predict($testSet);   // Tensor [N] of class indices
$score = $clf->score($testSet);     // float accuracy

// Feature importance
$fi = $clf->featureImportances();   // Tensor [D]

// Save / load
$clf->save('models/gbdt/');
$clf2 = GBDTClassifier::load('models/gbdt/');

// Cross-validation with any estimator
use Pml\CrossValidation\StratifiedKFold;

$cv     = new StratifiedKFold(k: 5);
$scores = $cv->score(new GBDTClassifier(), $dataset);
echo "Mean CV accuracy: " . array_sum($scores) / count($scores);

// Hyperparameter search
use Pml\Estimators\Meta\GridSearch;

$gs = new GridSearch(
    estimator: new GBDTClassifier(),
    grid: [
        'nEstimators' => [100, 200, 300],
        'maxDepth'    => [4, 6, 8],
        'lr'          => [0.05, 0.1],
    ],
    metric: 'accuracy',
    cv:     3,
);
$gs->train($dataset);
echo "Best params: "; print_r($gs->bestParams());