Scalable Hierarchical Multitask Learning in Sequence Biology
Christian Widmer, Jose Leiva-Murillo, Yasemin Altun and Gunnar Raetsch
In: Machine Learning for Computational Biology Workshop 2009, 11 Dec 2009, Whistler, Canada.
Multitask learning methods investigate the challenge of combining information
from several related problem domains. For a large family of multitask problems,
relationships between tasks can be described by a hierarchical structure. This is
particularly the case for many problems in Computational Biology, where differ-
ent tasks correspond to different organisms, whose relationship to each other is
deﬁned by phylogeny. In this work we describe a set of algorithms that com-
bine large-scale classiﬁcation methods with ideas from Domain Adaptation and
Multitask Learning. We introduce several methods of exploiting hierarchical task
relations for Multitask Learning. These algorithms are designed for large-scale
applications and scale to problems with a great number of training examples. The
performance of the presented methods is demonstrated on synthetic data, as well
as on splice-site data arising from a problem in genomic sequence analysis.