KNN on Iris

// Iris: Fisher, R. (1936). UCI. https://doi.org/10.24432/C56C76
// CC BY 4.0: https://creativecommons.org/licenses/by/4.0/
// Original UCI records retained; class names mapped to zero-based IDs.
// Full attribution and variant notes: datasets/iris/README.md.
import aner.dataset;
import aner.tensor;
import aner.ml;
import aner.metrics;

fn main() -> Unit {
    let flowers = dataset.iris();
    let split_seed = 2026;
    let partitions = dataset.split(flowers, 0.2, split_seed);
    let training = dataset.train(partitions);
    let testing = dataset.test(partitions);
    // Fit preprocessing only on training rows; reuse it unchanged for test rows.
    let scaler = dataset.fit_standardizer(dataset.features(training));
    let x_train = dataset.transform(scaler, dataset.features(training));
    let x_test = dataset.transform(scaler, dataset.features(testing));
    let test_labels = dataset.targets(testing);

    print("Dataset citation"); print(dataset.citation(flowers));
    print("Dataset license"); print(dataset.license(flowers));
    print("Dataset version"); print(dataset.version(flowers));
    print("Dataset SHA256"); print(dataset.sha256(flowers));
    print("Split seed"); print(split_seed);
    print("Training rows"); print(dataset.rows(training));
    print("Test rows"); print(dataset.rows(testing));

    // Choose k before looking at the held-out results.
    let k = 5;
    let model: KNNClassifier = ml.fit_knn(x_train, dataset.targets(training), k);
    let predictions = ml.predict(model, x_test);
    print("Neighbors"); print(k);
    print("Test accuracy"); print(metrics.accuracy(predictions, test_labels));
    let confusion = metrics.confusion_matrix(predictions, test_labels, dataset.classes(flowers));
    print("Confusion matrix: true rows, predicted columns; row-major counts");
    var row = 0;
    while row < tensor.rows(confusion) {
        var col = 0;
        while col < tensor.cols(confusion) {
            print(tensor.value(confusion, row, col));
            col = col + 1;
        }
        row = row + 1;
    }
    let class_ids = ml.classes(model);
    let votes = ml.predict_proba(model, x_test);
    print("First test row: class ID then neighbor vote fraction (not calibrated)");
    var c = 0;
    while c < tensor.rows(class_ids) {
        print(tensor.value(class_ids, c, 0));
        print(tensor.value(votes, 0, c));
        c = c + 1;
    }
    // Indices refer to x_train, not the original dataset; use row_ids to map back.
    let neighbors = ml.neighbor_indices(model, x_test);
    let distances = ml.neighbor_distances(model, x_test);
    let source_rows = dataset.row_ids(training);
    print("First test row: nearest training-local row index and standardized Euclidean distance");
    // Aner currently has no Float64-to-Int64 cast, so print the training-local
    // index separately; source_rows is the complete training-to-source mapping.
    print(tensor.value(neighbors, 0, 0));
    print(tensor.value(distances, 0, 0));
    print("Training-to-source mapping rows"); print(tensor.rows(source_rows));
}