Tip of the Day:January 24
sasCommunity Tip of the Day
Friedman proposed Regularized Discriminant Analysis (RDA) to overcome multicollinearity and other causes that make LDA/QDA ill-conditioned. The core idea is to regularize illy-conditioned within class covariance matrix with the pooled covariance matrix for QDA or to regularize illy-conditioned pooled covariance matrix with a diagonal matrix for LDA. For details about this algorithm, check the book: Elements of Statistical Learning, Chapter 4, section 3.1.
To implement RDA, we output sufficient statistics using OUTSTAT= in PROC DISCRIM and make appropriete changes to relavant statistics, then use the scoring functionality fo PROC DISCRIM to re-score the data with regularized covariance matrix. See link below for sample code on Regularized LDA.
Submitted By Liang Xie
Feel free to comment on this tip.