ML//evaluation//negative control

A negative control is a feature (or input) plausible enough to tempt the model and constructed so that it should carry no real signal: a trap a pipeline sets for itself. If feature selection gives it weight anyway, the pipeline has leakage, samples that are too small, or a model grabbing correlation-shaped noise.

A negative control is a feature (or input) plausible enough to tempt the model and constructed so that it should carry no real signal: a trap a pipeline sets for itself. If feature selection gives it weight anyway, the pipeline has leakage, samples that are too small, or a model grabbing correlation-shaped noise.

The idea is borrowed from experimental science, where it is the sample that must come back negative for the assay to be trusted. Passing it is one of the few ways a small pipeline earns trust.