Relief feature weights

Features that separate neighbors of different classes

A feature is useful when two nearby points that share a label agree on it, and two nearby points with different labels disagree on it. Relief scores every feature with exactly that idea, using one nearest same-class neighbor (the hit) and one nearest other-class neighbor (the miss) for each sample.

Implement relief_weights(X, y). It returns one weight per feature.

For each sample i:

1. Find the hit: the nearest other sample with the same label, by Euclidean distance on all features. Ties go to the smallest index.
2. Find the miss: the nearest sample with a different label, by the same rule.
3. For every feature j, add (|X[i, j] - X[miss, j]| - |X[i, j] - X[hit, j]|) / range_j to w[j], where range_j = max - min of feature j.

Divide the total by n.

Constraints

• X is 2-D with at least two rows, y has one label per row, otherwise raise ValueError.
• A feature with zero range contributes nothing. Use a divisor of 1 for it.
• If a sample has no hit or no miss (its class is the only one present), skip that sample but still divide by n.
• Do not modify the inputs.

Hints