ICML 2011 workshop on unsupervised and transfer learning

Supervised Dimensionality Reduction in the Unsupervised and Transfer Learning 2011 Competition

Yann-Aël Le Borgne (Tryan team, Fourth place in second ranking of phase 2 UTL challenge),
VUB, Belgium

This paper discusses preliminary results in the use of a supervised dimensionality reduction strategy to tackle the 2011 Unsupervised and Transfer Learning Competition. The basis of the strategy is to make use of the three subsets provided for each dataset in order to assign labels to examples. Using those labels, we rely on the partial least square technique to find a linear transformation that extract features of interest. We investigated two approaches. The first only uses three classes, assigned to the development, validation, and final subsets. The second additionally makes use of the transfer labels provided in the second phase of the competition. Our results show that the extracted features improve the classification accuracy upon the baseline results for four of the six datasets. Although these results are still preliminary, we believe that the proposed strategy provides a  flexible framework in which the performance could be greatly increased.