Definition
Plain language
A copy of an example with tiny changes, so it looks new but has the same answer.
As stated in the literature
Test rows with slightly perturbed feature values but identical targets to training rows; nearest-neighbor lookup then achieves near-perfect visible scores while exact-duplicate detection misses the leakage.
Also called: near-duplicates, near duplicate
Why it matters: Standard checks for identical rows sail right past it, so this kind of leakage can inflate scores without anyone noticing.
For example, a test row that lists a customer's age as 34.01 instead of 34 but carries the same answer looks brand new while being effectively a copy.
Heard on the show
“Sixty-four Searchers asked the same question would have produced sixty-four near-duplicate trajectories all confidently landing on Jesse Duroha.”Episode 051 — Why Parallel Sampling Plateaus, And What Evidence Graphs Do Instead