feature selection - MATLAB ReliefF Output -

the outputs of matlab feature selection algorithm relieff ranked , weights.

http://in.mathworks.com/help/stats/relieff.html

how 2 outputs related?

from docs:

ranked indices of columns in x ordered attribute importance, meaning ranked(1) index of important predictor. weight attribute weights ranging -1 1 large positive weights assigned important attributes.

so relieff doesn't give list of predictors important, gives list of weights give idea of how important predictors are. weight(4), example, weight of fourth predictor. weight(ranked) return ordered list of weights rank (highest lowest).

you note weight values predictors 4 & 3 in case relatively close. might indicate although particular set of data 4th predictor appears ranked first, 4th , 3rd predictors both important , given different set of data ranking change.

for example, fisheriris data set contains 150 measurements of 4 predictors. if take two-thirds of them can different result:

load fisheriris [ranked,weight] = relieff(meas(1:100,:),species(1:100),10) 

this gives us:

ranked = 3 4 2 1 weight = 0.1574 0.2265 0.5431 0.4981 

there 3 academic papers referenced in documentation on relieff. if want understand details of it's doing, should try getting hold of those.