Overfitting

/ˈoʊvərˌfɪtɪŋ/o·ver·fit·tingnoun
Statistical Learning Core

Definition

1.[in statistical learning] the failure mode in which a model absorbs noise particular to its training sample and therefore predicts unseen data worse than a simpler model would.

Validation error rose while training error kept falling — textbook overfitting.

Formal statement

E[(y - fhat(x))^2] = Bias[fhat]^2 + Var[fhat] + sigma^2

Overfitting is the regime where the variance term dominates the decomposition.

Etymology

From over- 'excessively' plus fit, in the statistical sense of fitting a curve to points; in use in the regression literature from the mid-twentieth century.

Synonyms

  • high variancesense 1 · Near

Antonyms

  • underfittingsense 1 · Gradable

See also

References