Model Selection Metrics


Metrics:

  • -> equivalent to SSE, RMSE. Focus on prediction performance
  • adj- -> evaluate both SSE and df (n-p), in MLR, adj- than
  • t test p-values(or F test p-values) -> focus on the statistical significance of predictors
  • AIC and BIC:
    • Akaike’s Information criterion(AIC):
      • AIC = (p: number of parameters, is the sample likelihood of the model):
        • 2p is a penalty on the number of predictors, adding more predictors might cause increase in AIC
        • AIC tends to choose model with large likelihood, without adding too many predictors
      • smaller ALC indicate better model
    • Bayesian Information criterion
      • BIC =
        • ln n is a much larger penalty on adding predictors to the model
    • Select model with the smallest AIC/BIC, they most of the time agree. if not, BIC picks shorter models, therefore more popular in MLR
  • Mallow’s Cp:
    • a good cp:
      • then choose one with smallest cp

Key words:


TAGS