PLS-EDA (Enhanced Discriminant Analysis using PLS)
PLS-EDA (PLS Enhanced Discriminant Analysis) was developed to separate categories more clearly than PLS-DA, and to extract key variables which contribute separation of categories. PLS-EDA is developed based on O-PLS-DA algorithm. In this chapter, let's see the differences between PLS-EDA and PLS-DA.
To test PLS-EDA, go back to "Appliance" sheet and click gMultibase_2015h and gOpen Formh. When Preparation dialog box appears, select gPLS-EDAh and click gDoneh button.
After "PLS-EDA" dialog box opens, click "Done" without any changes.
New worksheet "Multibase_PLSEDA" and "Multibase_Result" will be generated and loading and score plot with PC1, PC2 and PC3 will be displayed on Multibase_PLSEDA sheet as below.
You can see that the separation between groups become much clearer than PLS-DA. You can see "facsimile" contributes to the separation of Pennsylvania (red) and the other states.
Axis Rotation for Factor Matching
Understanding of underlying factors given by PCA, PLS-DA and PLS-EDA are sometimes difficult. In this case, rotating the axes helps you interpret factor meanings better. In this chapter, let's understand the effect of gAxis Rotation for Factor Matchingh.
Go back to "Multibase_Preparation" and click gMultibase_2015h and gOpen Formh in Add-ins menu. When Calculation dialog box appears, check gAxis Rotation for Factor Matchingh and click gDoneh button.
New worksheet "Multibase_PLSEDA" and "Multibase_Result" will be generated and the loading and the score plots with PC1, PC2 and PC3 will be displayed on Multibase_PLSEDA sheet as shown below.
gAxis Rotation for Factor Matchingh maximizes variable distribution on (or around) axes. In the case of Florida (Blue), Car Speakers, SD Cards, and Stereos & Components are distributed on left-hand side of PC1, and it means that those appliances are sold well in Florida.
2-3-1 Marunouchi Chiyoda Tokyo Japan