Spectral Feature Selection for Data Mining

Multi-Source Spectral Feature Selection 153

feature interactions, data can be filtered by the features associated with F, or

I:

X

F

= Π

F

F

(X) , X

I

= Π

F

I

(X) . (6.10)

Here F

F

and F

I

are the features related to F and I, respectively, and Π (·) is the

projection operator. Using the filtered data X

F

or X

I

, we can obtain a pairwise

sample similarity matrix S through any similarity measure. Since all features

in F

F

(or F

I

) are related to the feature functions (or feature interactions)

of interest, sample similarity matrix S should reflect the distribution under

the influence of the functions (or the interactions). In case that the functions

(or the interactions) are closely ...