outliers/

• IQROutlierDetector: Flags values beyond a multiple of the interquartile

range. Quartiles are themselves resistant to outliers, so the rule does not get dragged around by the points it is trying to find.

• ValueClipper: Caps values at explicit bounds, keeping the row and its

label while limiting how far the value can pull a fit.

Clipping preserves your sample size; dropping rows does not, and it discards labels along with them. Prefer clipping unless the extremes are known to be recording errors.

An outlier is not automatically noise. In fraud, fault and intrusion detection the rare rows are the signal — there, model them with anomaly rather than removing them here.