Choosing an estimator

Start from the task. Each row names the estimators to try first and links the guide page for that family. Every estimator follows the scikit-learn conventions: construct it with hyperparameters, then call fit, predict, predict_proba, transform or score. Inputs are float32 arrays.

Task	Try first	Also available	Guide
Classify rows of tabular data	RandomForestClassifier, GradientBoostingClassifier (two classes)	ExtraTreesClassifier, LogisticRegression, SVC, KNeighborsClassifier	Trees and gradient boosting
Classify into more than two classes with boosting	GradientBoosting with loss="MultiClass"	RandomForestClassifier	Trees and gradient boosting
Predict a number	GradientBoostingRegressor, RandomForestRegressor	Ridge, LinearRegression, SVR, KernelRidge	Trees and gradient boosting, Linear models
Rank items within groups	GradientBoosting with loss="YetiRank" or loss="PairLogit"	GradientBoosting with loss="QueryRMSE"	Trees and gradient boosting
Fit an interpretable linear model	Ridge, LogisticRegression	Lasso, ElasticNet, RidgeCV, LassoCV	Linear models
Predict with an uncertainty estimate	GaussianProcessRegressor	GaussianProcessClassifier, BayesianRidge	Gaussian processes
Group similar rows	KMeans	MiniBatchKMeans, GaussianMixture, DBSCAN, HDBSCAN	Clustering
Reduce the number of features	PCA	TruncatedSVD, IncrementalPCA, NMF, FastICA	Decomposition and embeddings
Draw a 2-D map of the data	UMAP	TSNE, Isomap, SpectralEmbedding	Decomposition and embeddings
Find the nearest rows exactly	NearestNeighbors	KNeighborsClassifier, KNeighborsRegressor	Neighbors and vector search
Search millions of vectors quickly	IVFIndex, CagraIndex	IVFPQIndex, IVFSQIndex	Neighbors and vector search
Flag unusual rows	IsolationForest	LocalOutlierFactor, OneClassSVM, EllipticEnvelope	Trees and gradient boosting
Forecast a series	ARIMA, ETS	AutoARIMA, Theta, ExponentialSmoothing	Time series
Split a series into trend and season	STL	ExponentialSmoothing	Time series
Scale, encode or impute features	StandardScaler, OneHotEncoder, SimpleImputer	MinMaxScaler, RobustScaler, KBinsDiscretizer	Preprocessing
Keep the most useful features	SelectKBest	VarianceThreshold, RFE	Preprocessing
Score a model with cross-validation	cross_val_score()	mojolearn.metrics	API reference
Build or run neural network layers	TransformerBlock, SmallMLPTrainer	SambaStack, Embedding	Neural networks
Train or run a language model	LanguageModelTrainer	BpeTokenizer	Language models

A first pass on tabular data

When you are unsure, fit one tree ensemble and one linear model and compare their scores on held-out rows:

import numpy as np
import mojolearn

rng = np.random.default_rng(0)
X = rng.random((10_000, 20), dtype=np.float32)
y = (X[:, 0] + X[:, 1] > 1.0).astype(np.int64)
X_train, X_test, y_train, y_test = X[:8_000], X[8_000:], y[:8_000], y[8_000:]

forest = mojolearn.RandomForestClassifier(n_estimators=100).fit(X_train, y_train)
linear = mojolearn.LogisticRegression().fit(X_train, y_train)
print(forest.score(X_test, y_test), linear.score(X_test, y_test))

Devices and numeric modes

In identical mode, mojolearn returns the same bits on Apple, NVIDIA and AMD GPUs. identical is the default, so every example in these guides runs in it without extra settings. On Apple GPUs you can choose fast instead, which is tuned for speed and does not promise identical bits:

mojolearn.set_numeric_mode("fast")

See Numeric modes and Devices and the CPU.