7.00.02 - Background - Aster Analytics

Teradata Aster® Analytics Foundation User GuideUpdate 2

Product
Aster Analytics
Release Number
7.00.02
Published
September 2017
Content Type
Programming Reference
User Guide
Publication ID
B700-1022-700K
Language
English (United States)
Last Update
2018-04-17

Regularization

Regularization is a technique for reducing overfitting and thus decreasing the variance of trained models. Generalized linear models (GLM) are fit by minimizing a loss function, such as the sum of squared errors. For example, given a predictor vector X ϵ p , a response variable Y ϵ , and N observation pairs, you can find model parameters β 0 and β with this formula:



These fits can be regularized by adding a penalty function P(β ) to the loss function being minimized. For example:



where λ controls the strength of the penalty function.

For logistic regression, the loss function is based on the log likelihood, as follows:



Three popular penalty functions are:
  • The sum of the absolute values of the model parameters:


    which is the L1 norm of the model parameters. This regularization technique, also called LASSO (for "least absolute shrinkage and selection operator"), was introduced by Robert Tibshirani in 1996. LASSO has the potential to shrink some parameters to zero; therefore, you can also use it for variable selection.

  • The sum of the squared values of the model parameters:


    which is the L2 norm of the model parameters. This regularization technique is also called ridge regression. With ridge regression, parameter values become smaller as λ increases, but never reach zero.

  • Elastic net regularization, which is a linear combination of L1 and L2 normalization:


The GLM2 function supports each of these regularization methods.

Coordinate Descent

GLM2 minimizes the loss function using coordinate descent, which is a multivariable minimization technique that minimizes the loss function along one β dimension at a time. For minimization problems, including the L1 norm, coordinate descent often performs better than the Fisher scoring method used by the GLM function.

References

  • Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1), 1 - 22.doi (GLM regularization paths article)

  • Tibshirani, R., Bien, J., Friedman, J., Hastie, T., Simon, N., Taylor, J. and Tibshirani, R. J. (2012), Strong rules for discarding predictors in lasso-type problems. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 74: 245–266. doi:10.1111/j.1467-9868.2011.01004.x