Out-of-Fold Target Encoding

Background

Naive target encoding leaks the target into the features. Out-of-fold (K-fold) target encoding fixes this: for each row, its encoding is computed only from rows outside its fold. Production tabular pipelines use this as their default category encoder.

Problem statement

Implement oof_target_encode(values, targets, n_folds=5, smoothing=10.0, seed=0). For each row, compute its smoothed target encoding using only rows from the other folds.

Input

• values - sequence of hashable category labels.
• targets - parallel sequence of numeric targets.
• n_folds - int >= 2.
• smoothing - float >= 0.
• seed - RNG seed for fold assignment.

Output

• list[float] of length $n$, the per-row OOF encoding.

Examples

Input: 4 rows of category "A" with targets [1,1,1,0], n_folds=2
Output: row in fold 0 uses fold 1 stats; row in fold 1 uses fold 0 stats

Constraints

• Random fold assignment via the seed.
• For categories not present in the out-of-fold partition, use the global prior.

Notes

• Why OOF. Without it, training accuracy is inflated and the model overfits to its own encoding.