Model Selection

Utilities for splitting data into train/test sets and evaluating models via cross-validation.

train_test_split

Splits the data arrays $X$ and $y$ into random train and test subsets. Given $n$ samples and test fraction $f$:

$$n_{\text{test}} = \lfloor n \cdot f \rfloor, \qquad n_{\text{train}} = n - n_{\text{test}}$$

When shuffle = true (default), the data is randomly permuted before splitting using the Fisher–Yates shuffle with the given random_state seed. This ensures reproducibility.

#include <Skigen/ModelSelection>

auto [X_train, X_test, y_train, y_test] = Skigen::train_test_split(
    X, y,
    /*test_size=*/0.25,
    /*random_state=*/42,
    /*shuffle=*/true
);

Parameter	Default	Description
test_size	0.25	Fraction of data reserved for testing
random_state	42	Random seed for reproducibility
shuffle	true	Whether to shuffle before splitting

cross_val_score

Evaluates a model using $K$-fold cross-validation. The data is split into $K$ non-overlapping folds. For each fold $k$:

1. Train the model on all folds except fold $k$.
2. Score the model on fold $k$.

The function returns a vector of $K$ scores. The mean score estimates the model's generalization performance.

#include <Skigen/ModelSelection>

Skigen::LinearRegression model;
auto scores = Skigen::cross_val_score(model, X, y, /*cv=*/5);
std::cout << "Mean R²: " << scores.mean() << "\n";
std::cout << "Std:     " << std::sqrt((scores.array() - scores.mean()).square().mean()) << "\n";

Parameter	Default	Description
cv	5	Number of folds $K$
shuffle	true	Whether to shuffle before folding
random_state	42	Random seed for reproducibility

Mirrors sklearn.model_selection.train_test_split and cross_val_score.