Spurious Patterns and Proxy Features

Prediction can exploit the wrong reason

A spurious correlation is an association that does not reflect a stable mechanism and may disappear when conditions change. A proxy feature indirectly represents another quantity, such as hospital ID proxying for local treatment policy or image background proxying for class labels.

Models are rewarded for prediction, not for discovering the reason humans intended. If every training photo of wolves contains snow and dog photos contain grass, a classifier may learn background color rather than anatomy. It can score well on a similarly collected test set and fail on a wolf photographed in summer.

training shortcut: snow -> wolf
intended concept: body shape -> wolf
stress test: wolf on grass exposes the shortcut

Analogy: A student who recognizes the answer key by a printing smudge can ace the practice test without learning the subject. Change the printer and the performance vanishes.

Tip: Use slice tests, changed environments, counterexamples, and domain review to probe whether a model depends on a stable signal or a collection artifact.