Affiliations: ISISTAN Research Institute, Unicen University, Campus Universitario, Tandil (B7001BBO), Argentina
Correspondence:
[*]
Corresponding author: Juan Manuel Rodriguez, ISISTAN Research Institute, Unicen University, Campus Universitario, Tandil (B7001BBO), Argentina. Tel.: +54 249 4385682 ext. 35; Fax: +54 249 4385681; E-mail:juanmanuel.rodriguez@isistan.unicen.edu.ar
Abstract: Large scale multi-label learning, i.e. the problem of determining
the associated set of labels for an instance, is gaining relevance
in recent years due to the emergence of several real-world applications.
Most notably, the exponential growth of the Social Web where a resource
can be labeled by millions of users using one or more tags, i.e. a
resource can be associated to several labels at the same time. A well-known
approach for multi-label classification is the Binary Relevance (BR)
algorithm which trains a binary classifier for each label independently.
However, the serial implementation of BR is not suitable for medium
or large datasets due to the time and computational resources required
for training. For example, training classifiers for mid-size datasets
using MULAN implementation of BR might take several weeks. This paper
discusses a parallel implementation of the MULAN BR technique that
harnesses the computational power of nowadays multi-core processors.
Our implementation presents a speed-up in the training phase of up
to 12 times when compared to the original MULAN implementation. In
addition, the cross-validation technique of MULAN had huge RAM requirements,
making it unusable with large datasets. Therefore, we have overcome
this limitation by using compact data structures and taking advantage
of disk caching. We have also compared our implementation against
scikit-learn, a popular tool for data mining and data analysis, showing
significant improvements in speed-up.
