sample-weighting

Weight samples by chunk behavior — leave-chunk-out influence over contiguous/interleaved/feature-cluster partitions; downweight regime-broken chunks. Held-out + null validated (honest about the overfit-the-tuned-split trap). numpy-only.

Jointly learn sample + feature weights to minimize the train-CV gap (stability / anti-overfit) — cross-fold coefficient stability + local consensus, held-out + null validated, transparent. numpy-only.

Fuse recency and residual-stability sample weights; requires fused vector to beat uniform, best single signal, and permutation null. numpy-only.

Learn a shallow, interpretable sample-weight map over driver features (window/neighbor/density/residual) — catches INTERACTION structure no single driver does. Must beat uniform, the best single driver, AND a null. numpy-only.

Downweight training samples that distort the fit, via CLOSED-FORM ridge influence (leverage / LOO residual / Cook's distance) — analytic, no resampling. Held-out + null validated. numpy-only.

Recency sample weighting for regime drift; selects recent-row emphasis on train only, then validates against held-out and permutation null. numpy-only.

Correct for redundancy in tabular/time-series ML — downweight dense / near-duplicate rows (effective-sample-size correction) so a fit is not dominated by autocorrelated repeats. Held-out + null validated. numpy-only.

Search periodic sample-weight patterns (sine/comb/sawtooth) that lift out-of-period transfer for time-ordered ML; bootstrap to robust per-record weights. numpy-only.

Per-sample weights from bootstrap stability — downweight rows whose out-of-bag residual is persistently large/unstable across resamples. Held-out + null validated. numpy-only.

Non-naive sample weighting from feature-space kNN behavior — downweight rows in isolated or inconsistent neighborhoods (local outlierness/label-dispersion/density). Held-out + null validated. numpy-only.