Imbalanced-data resampling

SMOTE, random over/under-sampling and balanced class weights for skewed datasets.

imbalance rebalances classification datasets where one class dominates — zero Python imports — replacing imbalanced-learn (SMOTE, RandomOverSampler, RandomUnderSampler) and sklearn.utils.class_weight.compute_class_weight("balanced"). It covers synthetic minority oversampling by interpolation between nearest minority neighbors, seeded random over/under-sampling to the majority/minority count, and inverse-frequency class weights.

The module

Module	What it does
imbalance	SMOTE (1-D and n-D), seeded random over/under-sampling, and balanced inverse-frequency class weights for skewed classification datasets, fully deterministic per seed

Functions

Function	What it does
smote_1d(x, labels, minority_label, k_neighbors, n_synthetic, seed)	SMOTE for a single feature: synthetic minority values, each on the interval between a minority value and one of its k nearest minority neighbors.
smote_nd(X, n_features, labels, minority_label, k_neighbors, n_synthetic, seed)	SMOTE for d features (row-major flat X): each synthetic row is x_b + gap * (x_nb − x_b) with gap ~ U[0, 1), so it lies on the segment between two minority rows.
random_oversample_indices(labels, seed)	Indices resampled so every class reaches the max class count: originals in order, then each class's deficit filled by sampling with replacement.
random_undersample_indices(labels, seed)	Keeps min-class-count samples per class, chosen without replacement (partial Fisher–Yates), returned in ascending order.
class_weights(labels)	Balanced inverse-frequency weights per class: weight_c = n / (n_classes * count_c) — sklearn's "balanced" mode.

Usage guidance

• Deterministic for a given seed. All samplers draw from a seeded xorshift64 RNG: same seed → identical output, different seed → different output (both test-verified). Already-balanced labels return the identity [0, n); empty labels return empty.

• Infeasible configs return empty, never trap. Fewer than 2 minority rows, k_neighbors < 1, n_synthetic < 1, n_features < 1, or a mis-sized X all yield an empty result; k_neighbors larger than n_minority − 1 is clamped.

• Cost. Nearest neighbors are computed lazily once per distinct base row and cached, so worst case is O(min(n_synthetic, n_min) · n_min · n_features). SMOTE is offline training-data prep (never in the live decision loop); for typical minority classes (n_min ≤ ~1k) it is sub-millisecond.

Verification

Geometry is test-pinned: with minority values {0, 10, 20} and majority values {100, 101}, all 16 synthetic values stay inside the minority hull [0, 20] — majority rows never leak into the interpolation. In 2-D with minority rows A=(0,0), B=(4,0), C=(4,3) (squared distances AB=16, AC=25, BC=9), every k=1 synthetic row lies on segment AB or BC — never AC — matching the exact 1-NN structure, and k=2 rows are always convex combinations of some minority pair. class_weights([0,0,0,1]) returns [2/3, 2] exact to 1e-12 with sample-weight mass preserved (sum count_c · w_c == n = 4), matching sklearn's "balanced" formula; oversampling [0,0,0,1] gives exactly [0,1,2,3,3,3], and a 3-class input balances every class to the max count of 3.

Next: promote what you trained with statistical model comparison, or browse the engine catalog.