pátek 6. března 2015

How to deal with overfitting

1) Measure it (with cross-validation...)
2) Decrease it
  - Get more data
  - Decrease the size of the hypotheses space (for example decrease the degree of polynomial in regression, limit the decision tree size, assume attribute independence in Bayes...)
  - Introduce bias (for example L1 or L2 regularization in regression, ensembles of different classifiers, operators background knowledge,...)

Žádné komentáře: