The HPREDUCE Procedure

This section compares the HPREDUCE procedure with the FACTOR, PRINCOMP, GLMSELECT, and DISCRIM procedures in SAS/STAT software.

When PROC HPREDUCE performs unsupervised variable selection, it conducts variance analysis and reduces dimensionality by forward selection of the variables that contribute the most to the overall data variance. The output lists the variables in order of their contribution to data variance and can be used directly for reporting or for selecting variables for model building procedures. In contrast, principal component analysis (PCA) conducts a variance analysis and then projects the data space to an orthogonal set of axes by a linear combination of the original variables. These new principal components best explain the data variance and can be used as input to model building procedures. In either case, the number of inputs to the modeling procedure has been reduced from the original set. PCA can be done through the SAS/STAT FACTOR and PRINCOMP procedures. The primary difference between PCA and PROC HPREDUCE is that PCA generates new variables, while PROC HPREDUCE reduces data dimensionality by selecting a subset of the original variables. This feature of PROC HPREDUCE is beneficial in applications where retaining the original variables is important for model exploration and interpretation.

When PROC HPREDUCE performs supervised variable selection, it conducts variance analysis and reduces dimensionality by forward selection of the variables that contribute the most to explaining the overall variance of the response variables (targets). The output lists the variables in order of their contribution to explaining response variance. The output can be used directly for reporting or for selecting variables for model building procedures. When PROC HPREDUCE is used to perform supervised variable selection, it most resembles the GLMSELECT procedure. However, PROC HPREDUCE allows multiple response variables, which is not supported by PROC GLMSELECT. When the response variable is a classification variable and its levelization is done in a special format, PROC HPREDUCE conducts variance analysis in the same way as linear discriminant analysis (LDA) does. LDA can be done through the SAS/STAT DISCRIM procedure. Like PCA, LDA generates new variables by linearly combining all original variables, while PROC HPREDUCE reduces data dimensionality by selecting a subset of the original variables.