2
1

Hi,

can someone recommend some opensource Python code for multi-task supervised learning? e.g. Multi-Task SVM or Multi-Task sparse multinomial regression or whatever comes to your mind.

Thanks

asked Jan 03 '11 at 17:18

Alexandre%20Gramfort's gravatar image

Alexandre Gramfort
91237

to give more details I have 2 objectives :

  • benchmark a solution for predicting multiple binary outputs jointly.

  • see what's available out there in order to add this feature to scikit-learn

do not hesitate to promote your own work !

I am also interested by answers like : "I would try with [?] (it's implemented there in Matlab)" or "I would implement [?] it is fairly straightforward"

(Jan 04 '11 at 21:40) Alexandre Gramfort

Do you have some reference papers in mind?

(Jan 06 '11 at 04:08) ogrisel

sure.

Caruana. Multitask Learning: A Knowledge-Based Source of Inductive Bias. (1995)

Ben-David et Schuller. Exploiting task relatedness for multiple task learning. Learning Theory and Kernel Machines (2003) pp. 567-580

Ando et Zhang. A framework for learning predictive structures from multiple tasks and unlabeled data. The Journal of Machine Learning Research (2005) vol. 6 pp. 1817-1853

Evgeniou et Pontil. Regularized multi-task learning. Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining (2004) pp. 109-117

This interest follows a recent reading at NIPS 2010:

Decoding Ipsilateral Finger Movements from ECoG Signals in Humans Y. Liu, M. Sharma, C. Gaona, J. Breshears, j. Roland, z. Freudenburg, K. Weinberger, E. Leuthardt

do not hesitate to add more if I miss some key contributions to this problem.

(Jan 06 '11 at 10:31) Alexandre Gramfort

Minor comment: there is a more complete article by Caruana in the journal 'Machine Learning'.

(Jan 27 '11 at 15:48) Art Munson

One Answer:

Alexandre,

my NUT project contains a (python) implementation of Ando & Zhangs ASO algorithm. The algorithm is pretty straight-forward: learn multiple linear predictors independently (via multiprocessing or Hadoop), then run SVD on the predictor matrix.

best, Peter

answered Jul 18 '11 at 11:00

Peter%20Prettenhofer's gravatar image

Peter Prettenhofer
35579

Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.