Linear Algorithms for Online Multitask Classification
Abstract
We design and analyze interacting online algorithms for multi-task classification that perform better than independent learners whenever the tasks are related in a certain sense. We formalize task relatedness in different ways, and derive formal guarantees on the performance advantage provided by interaction. Our online analysis gives new stimulating insights into previously known co-regularization techniques, such as the multi-task kernels and the margin correlation analysis for multi-view learning. In the last part we apply our approach to spectral co-regularization: we introduce a natural matrix extension of the quasiadditive algorithm for classification and prove bounds depending on certain unitarily invariant norms of the matrix of task coefficients.
Cite
Text
Cavallanti et al. "Linear Algorithms for Online Multitask Classification." Annual Conference on Computational Learning Theory, 2008. doi:10.5555/1756006.1953026Markdown
[Cavallanti et al. "Linear Algorithms for Online Multitask Classification." Annual Conference on Computational Learning Theory, 2008.](https://mlanthology.org/colt/2008/cavallanti2008colt-linear/) doi:10.5555/1756006.1953026BibTeX
@inproceedings{cavallanti2008colt-linear,
title = {{Linear Algorithms for Online Multitask Classification}},
author = {Cavallanti, Giovanni and Cesa-Bianchi, Nicolò and Gentile, Claudio},
booktitle = {Annual Conference on Computational Learning Theory},
year = {2008},
pages = {251-262},
doi = {10.5555/1756006.1953026},
url = {https://mlanthology.org/colt/2008/cavallanti2008colt-linear/}
}