src/smote.py

# minimal SMOTE implementation (no imbalanced-learn dependency)
import numpy as np


def smote(X, y, k=5, random_state=42):
    rng = np.random.default_rng(random_state)

    classes, counts = np.unique(y, return_counts=True)
    minority_class  = classes[np.argmin(counts)]
    majority_class  = classes[np.argmax(counts)]
    n_majority      = counts[np.argmax(counts)]
    n_minority      = counts[np.argmin(counts)]
    n_synthetic     = n_majority - n_minority

    X_min = X[y == minority_class]

    # pairwise euclidean distances among minority samples
    diffs    = X_min[:, None, :] - X_min[None, :, :]
    sq_dists = (diffs ** 2).sum(axis=2)
    np.fill_diagonal(sq_dists, np.inf)
    k      = min(k, n_minority - 1)
    nn_idx = np.argsort(sq_dists, axis=1)[:, :k]

    synthetic = np.empty((n_synthetic, X.shape[1]))
    for i in range(n_synthetic):
        base      = rng.integers(0, n_minority)
        neighbour = nn_idx[base, rng.integers(0, k)]
        lam       = rng.uniform(0, 1)
        synthetic[i] = X_min[base] + lam * (X_min[neighbour] - X_min[base])

    X_res = np.vstack([X, synthetic])
    y_res = np.concatenate([y, np.full(n_synthetic, minority_class)])

    # shuffle so minority examples aren't all at the end
    idx = rng.permutation(len(y_res))
    return X_res[idx], y_res[idx]