The multi-label classification problem has generated significant interest in
cent years. However, existing approaches do not adequately address two key
allenges: (a) the ability to tackle problems with a large number (say
llions) of labels, and (b) the ability to handle data with missing labels. In
is paper, we directly address both these problems by studying the multi-label
oblem in a generic empirical risk minimization (ERM) framework. Our
amework, despite being simple, is surprisingly able to encompass several
cent label-compression based methods which can be derived as special cases of
r method. To optimize the ERM problem, we develop techniques that exploit the
ructure of specific loss functions - such as the squared loss function - to
fer efficient algorithms. We further show that our learning framework admits
rmal excess risk bounds even in the presence of missing labels. Our risk
unds are tight and demonstrate better generalization performance for low-rank
omoting trace-norm regularization when compared to (rank insensitive)
obenius norm regularization. Finally, we present extensive empirical results
a variety of benchmark datasets and show that our methods perform
gnificantly better than existing label compression based methods and can
ale up to very large datasets such as the Wikipedia dataset.

@misc{yu2013largescale, author = {Yu, Hsiang-Fu and Jain, Prateek and Kar, Purushottam and Dhillon, Inderjit S.}, title = {Large-scale Multi-label Learning with Missing Labels}, year = {2013}, note = {cite arxiv:1307.5101}, url = {http://arxiv.org/abs/1307.5101}, keywords = {classification, kallimachos, label, large, learning, multi}, abstract = {The multi-label classification problem has generated significant interest inrecent years. However, existing approaches do not adequately address two keychallenges: (a) the ability to tackle problems with a large number (saymillions) of labels, and (b) the ability to handle data with missing labels. Inthis paper, we directly address both these problems by studying the multi-labelproblem in a generic empirical risk minimization (ERM) framework. Ourframework, despite being simple, is surprisingly able to encompass severalrecent label-compression based methods which can be derived as special cases ofour method. To optimize the ERM problem, we develop techniques that exploit thestructure of specific loss functions - such as the squared loss function - tooffer efficient algorithms. We further show that our learning framework admitsformal excess risk bounds even in the presence of missing labels. Our riskbounds are tight and demonstrate better generalization performance for low-rankpromoting trace-norm regularization when compared to (rank insensitive)Frobenius norm regularization. Finally, we present extensive empirical resultson a variety of benchmark datasets and show that our methods performsignificantly better than existing label compression based methods and canscale up to very large datasets such as the Wikipedia dataset.} }

%0 = misc %A = Yu, Hsiang-Fu and Jain, Prateek and Kar, Purushottam and Dhillon, Inderjit S. %B = } %C = %D = 2013 %I = %T = Large-scale Multi-label Learning with Missing Labels} %U = http://arxiv.org/abs/1307.5101

@inproceedings{noauthororeditor, author = {Mirowski, Piotr and Ranzato, Marc'Aurelio and LeCun, Yann}, title = {Dynamic Auto-Encoders for Semantic Indexing}, editor = {of the NIPS 2010 Workshop on Deep Learning, Proceedings}, year = {2010}, url = {http://yann.lecun.com/exdb/publis/pdf/mirowski-nipsdl-10.pdf}, keywords = {deep, kallimachos, lda, learning, model, toread} }

%0 = inproceedings %A = Mirowski, Piotr and Ranzato, Marc'Aurelio and LeCun, Yann %D = 2010 %T = Dynamic Auto-Encoders for Semantic Indexing %U = http://yann.lecun.com/exdb/publis/pdf/mirowski-nipsdl-10.pdf