Unified Training Interface

smile.model.Model provides one entry point for tabular classification and regression. Choose an algorithm name, a Formula, data, and a Properties bag of hyperparameters; the result is a ClassificationModel or RegressionModel record containing the predictor, metrics, schema, formula, and metadata.

Use the focused classification and regression guides for algorithm theory and complete Options references.

Quick Start

This Random Forest example trains on Iris with a small Properties configuration. Passing null for the test frame still returns the rich model record.

import java.util.Properties;
import smile.datasets.Iris;
import smile.model.Model;

var iris = new Iris();
var params = new Properties();
params.setProperty("smile.random_forest.trees", "200");

var model = Model.classification(
    "random-forest", iris.formula(), iris.data(), null, params);
System.out.println(model.algorithm());
System.out.println(model.train().accuracy());

random-forest
(training accuracy)

Core Concepts

algorithm()
The exact dispatch name, such as random-forest.
schema()
Predictor schema with the response column removed.
formula()
The formula bound during training.
tags()
Mutable metadata and a cloned snapshot of training properties.

Formula.lhs("y") selects y as the response and uses all remaining columns as predictors. The supplied Properties is cloned when the record is created: later changes to the original do not change model.tags().

import java.util.Properties;
import smile.datasets.Iris;
import smile.model.Model;

var iris = new Iris();
var params = new Properties();
var model = Model.classification(
    "cart", iris.formula(), iris.data(), null, params);
params.setProperty("after.training", "not copied");

System.out.println(model.formula());
System.out.println(model.schema());
System.out.println(model.tags().containsKey("after.training"));

(formula)
(predictor schema)
false

Training Classification Models

The five-argument overload accepts an optional held-out test frame. The full overload additionally controls repeated cross-validation and whether fold predictors become the final ensemble.

classification(algorithm, formula, data, params)
Returns the underlying DataFrameClassifier; no metric record.
classification(algorithm, formula, data, test, params)
Returns ClassificationModel; test may be null.
classification(algorithm, formula, data, test, params, kfold, round, ensemble)
Adds repeated stratified CV and optional fold-model ensembling.
import java.util.Properties;
import smile.datasets.ImageSegmentation;
import smile.model.Model;

var data = new ImageSegmentation();
var plain = Model.classification(
    "random-forest", data.formula(), data.train(), null, new Properties());
var tested = Model.classification(
    "random-forest", data.formula(), data.train(), data.test(), new Properties());
var validated = Model.classification(
    "random-forest", data.formula(), data.train(), data.test(),
    new Properties(), 5, 1, false);
System.out.println(tested.test().accuracy());
System.out.println(validated.validation().accuracy());

(held-out accuracy)
(average CV accuracy)

Training Regression Models

Regression has the same three overload shapes. The four-argument form returns DataFrameRegression; the forms with a test argument return RegressionModel.

import java.util.Properties;
import smile.datasets.Abalone;
import smile.model.Model;

var abalone = new Abalone();
var model = Model.regression(
    "ols", abalone.formula(), abalone.train(), abalone.test(),
    new Properties());
System.out.println(model.test().rmse());

var cv = Model.regression(
    "ols", abalone.formula(), abalone.train(), abalone.test(),
    new Properties(), 5, 1, false);
System.out.println(cv.validation().r2());

(held-out RMSE)
(average CV R²)

Cross-Validation and Ensembles

Set kfold >= 2 to run round × kfold folds. Classification uses stratified folds; regression uses regression folds. validation() stores the average metrics. With ensemble=false, SMILE retrains on all training rows; with ensemble=true, the fold models are combined. When kfold < 2, validation is null and the flag is ignored.

For choosing folds and comparing candidates, see Model Validation. For searching property configurations, see Hyperparameter Optimization.

Evaluating Models

The records expose train(), validation(), and test(). Training metrics are always populated; validation and test metrics are nullable because their data is optional. Classification metrics include accuracy, errors, precision, recall, F1, and binary AUC where applicable. Regression metrics include RMSE, MAE, and R².

import java.util.Properties;
import smile.datasets.ImageSegmentation;
import smile.model.Model;

var data = new ImageSegmentation();
var model = Model.classification(
    "cart", data.formula(), data.train(), data.test(),
    new Properties(), 5, 1, false);
System.out.printf("train %.3f%n", model.train().accuracy());
System.out.printf("cv %.3f%n", model.validation().accuracy());
System.out.printf("test %.3f%n", model.test().accuracy());

train (accuracy)
cv (accuracy)
test (accuracy)

Making Predictions

Both records implement single-row predict(Tuple). Classification also accepts a posterior-probability array and exposes numClasses(). For batches, call classifier() or regression() and predict a DataFrame.

import java.util.Properties;
import smile.datasets.Iris;
import smile.model.Model;

var iris = new Iris();
var model = Model.classification(
    "random-forest", iris.formula(), iris.data(), null, new Properties());
int one = model.predict(iris.data().get(0));
double[] posterior = new double[model.numClasses()];
model.predict(iris.data().get(0), posterior);
int[] batch = model.classifier().predict(iris.data());
System.out.println(one + " / " + batch.length);

(predicted class) / 150

Metadata Tags

Tags are strings stored in the record's Properties. Model.ID and Model.VERSION are conventional keys; custom provenance keys are welcome. Read them with getTag(key) or getTag(key, defaultValue).

import java.util.Properties;
import smile.datasets.Iris;
import smile.model.Model;

var iris = new Iris();
var model = Model.classification(
    "cart", iris.formula(), iris.data(), null, new Properties());
model.setTag(Model.ID, "iris-cart");
model.setTag(Model.VERSION, "1.0");
model.setTag("dataset", "iris");
System.out.println(model.getTag(Model.ID));
System.out.println(model.getTag("environment", "development"));

iris-cart
development

Algorithm Reference

Names below are the exact strings accepted by the Model switch cases. Prefixes identify the most useful property families, not every available key.

Classification

NameAlgorithmKey property prefix
random-forestRandom Forestsmile.random_forest.*
gradient-boostGradient Boosted Treessmile.gradient_boost.*
ada-boostAdaBoostsmile.adaboost.*
cartDecision Treesmile.cart.*
logisticLogistic Regressionsmile.logistic.*
fisher, lda, qdaDiscriminant analysisnone
rdaRegularized Discriminant Analysissmile.rda.*
mlpMultilayer Perceptronsmile.mlp.*
svmSupport Vector Machinesmile.svm.*
rbfRBF Networkalgorithm defaults

Regression

NameAlgorithmKey property prefix
random-forestRandom Forestsmile.random_forest.*
gradient-boostGradient Boosted Treessmile.gradient_boost.*
cartRegression Treesmile.cart.*
olsOrdinary Least Squaressmile.ols.*
lassoLASSOsmile.lasso.*
elastic-netElastic Netsmile.elastic_net.*
ridgeRidge Regressionsmile.ridge.*
gaussian-processGaussian Processsmile.gaussian_process.*
mlpMultilayer Perceptronsmile.mlp.*
svmSupport Vector Regressionsmile.svm.*
rbfRBF Networkalgorithm defaults

Common examples include smile.random_forest.trees, smile.gradient_boost.shrinkage, smile.cart.max_depth, smile.svm.C, and smile.ols.method. Use Hyperparameter Optimization to generate candidate Properties configurations.

Kernel Specification Strings

SVM and Gaussian Process properties parse compact kernel strings. Supported forms include linear, Gaussian(sigma), Laplacian(sigma), Polynomial(degree,scale,offset), ThinPlateSpline(sigma), Pearson(omega,nu), Hyperbolic(scale,offset), Hellinger, SparsLinear, and BinarySparseLinear.

import java.util.Properties;
import smile.datasets.Iris;
import smile.model.Model;

var iris = new Iris();
var params = new Properties();
params.setProperty("smile.svm.kernel", "Gaussian(1.0)");
params.setProperty("smile.svm.C", "10");
params.setProperty("smile.svm.type", "ovo");
var svm = Model.classification(
    "svm", iris.formula(), iris.data(), null, params);
System.out.println(svm.train().accuracy());

(training accuracy)

Serialization

Both model records implement Serializable. smile.io.Write.object writes a record and smile.io.Read.object restores it. Keep serialized models with their compatible SMILE version; Java serialization is not a language-neutral interchange format.

import java.nio.file.Path;
import java.util.Properties;
import smile.datasets.Iris;
import smile.io.Read;
import smile.io.Write;
import smile.model.ClassificationModel;
import smile.model.Model;

var iris = new Iris();
var model = Model.classification(
    "cart", iris.formula(), iris.data(), null, new Properties());
Path path = Write.object(model);
var restored = (ClassificationModel) Read.object(path);
System.out.println(restored.predict(iris.data().get(0)));

(predicted class from restored model)

Complete End-to-End Example

This held-out Image Segmentation run trains, evaluates, tags, and uses one Random Forest model. The test frame is never used for fitting.

import java.util.Properties;
import smile.datasets.ImageSegmentation;
import smile.model.Model;

var data = new ImageSegmentation();
var params = new Properties();
params.setProperty("smile.random_forest.trees", "200");
params.setProperty("smile.random_forest.max_nodes", "100");

var model = Model.classification(
    "random-forest", data.formula(), data.train(), data.test(), params);
model.setTag(Model.ID, "segmentation-rf");
model.setTag(Model.VERSION, "1.0");

System.out.printf("train accuracy = %.3f%n", model.train().accuracy());
System.out.printf("test accuracy  = %.3f%n", model.test().accuracy());
int prediction = model.predict(data.test().get(0));
System.out.println("first prediction = " + prediction);

train accuracy = (value)
test accuracy = (value)
first prediction = (class)