Clasificación multi-etiqueta utilizando computación distribuida

Multi-label classification techniques have been developed for problems where objects can be associated to several disjoint labels, such as the scientific topics covered by a paper. However, these techniques tend to be computationally complex, which makes it difficult to use them in practice. Therefore, they might be unsuitable for large problems. This paper presents an approach to accelerate a well-know multi-label classification technique, called Binary Relevance, by using small computational clusters. In this classification technique, the training times grow linearly with the number of labels. In particular, this work aims at reducing the times required for training a Binary Relevance classifier. This approach was tested using 7 data-sets with 81 associated labels and more than a quarter million training instances. Experimental results shown a linear increment on the speed-up when computational nodes are added to the cluster.