FORMS OF PARTIAL LEAST SQUARE
• PLS1, PLS2: rank-one approximation of X,Y with a score vector t and vector of loadings p, q
– X→X−tpT ; Y→Y−tcT
– mutually orthogonal score vectors ti , i = 1, . . . , m
– 1st SVi+1 ≥ 2nd SVi → select one score vector at a time
• PLS Mode A: rank-one approximation of X,Y with score vectors t,u and vector of loadings p,q
– X→X−tpT ; Y→Y−uqT
– mutually orthogonal score vectors ti, ui , i = 1, . . . , m
• PLS-SB: SVD of YT X = AΣBT
– YTX→YTX−σabT
– mutually orthogonal weight vectors ai,bi
– generally not orthogonal score vectors ci = Xai , di = Ybi
• SIMPLS :(de Jong’93)
– avoids deflation of X; i.e. finds weight vectors w ̃i
s u c h t h a t T ̃ = X 0 W ̃
– SVD of XT0 Y0 + constraint of mutually orthogonal ̃ti
– sequence of SVD problems P ̃⊥i XT0 Y0
P ̃⊥i an orthogonal projector onto P ̃i = [p ̃1,…,p ̃i]
̃ ̃ ̃
where p ̃i = XT0 ti/(tTi ti) are loadings vectors
– same as PLS1 but differs for PLS2
• Hinkel & Rayens’98-00; Frank & Friedman’93: – constraint maximization of covariance
SUMBER : CITESEERX