دانلود Semisupervised Feature Selection Based on Relevance and Redundancy Criteria

Abstract

Feature   selection   aims   to   gain   relevant   features for  improved  classification  performance  and  remove  redundant features  for  reduced  computational  cost.  How  to  balance  these two  factors  is  a  problem  especially  when  the  categorical  labels are  costly  to  obtain.  In  this  paper,  we  address  this  problem using   semisupervised   learning   method   and   propose   a   max-relevance   and   min-redundancy  criterion   based   on   Pearson’s correlation (RRPC) coefficient. This new method uses the incremental  search  technique  to  select  optimal  feature  subsets.  The new  selected  features  have  strong  relevance  to  the  labels  in supervised manner, and avoid redundancy to the selected feature subsets  under   unsupervised   constraints.   Comparative   studies are  performed  on  binary  data  and  multicategory  data  from benchmark  data  sets.  The  results  show  that  the  RRPC  can achieve  a  good  balance  between  relevance  and  redundancy  in semisupervised  feature  selection.  We  also  compare  the  RRPC with classic supervised feature selection criteria (such as mRMR and  Fisher  score),  unsupervised  feature  selection  criteria  (such as Laplacian score), and semisupervised feature selection criteria(such  as   sSelect  and   locality   sensitive).   Experimental   results demonstrate the effectiveness  of  our method.