Unformatted text preview: Multilabel Prediction via C ompressed Sensing
Daniel Hsu (UCSD), Sham M. Kakade (TTI-C), John Langford (Yahoo!), Tong Zhang (Rutgers) Goal: multilabel prediction; # of possible labels d is very large. : {car7 robot, lazer eyes} E Rd (year : 17 yI‘ObOt : 1: ylazer eyes Z 1)
What we exploit: output sparsity, i.e. E[y|a:] is k—sparse, k << d. \
r 827“ Training and PI‘EWA e Rde, m = 0(klog d)
a {(51% y )} '—> {(9371431)} '—> gzx—mum Training data ’\ Compressed training data Predictor of compressed labels Compressed Sparse
sensing reconstruction a: H §($)6Rm ¥> yeRd Test point Predicted compressed label Reconstructed sparse label a ...
View
Full Document
- Fall '09
- JilinWang
- Computer Science
-
Click to edit the document details