Exploration of usability of PLSR for implementation in the RENT feature selection method

Abstract

RENT (Repeated Elastic Net Technique) is a feature selection technique developed for binary classification and regression tasks. But most real life cases are multi-class.

RENT is not currently capable of handling multi-class classification or regression problems. Our thesis is an attempt to extend RENT to handle multi-class problems. To this end we have explored the PLSR algorithm to study if it is a good option for multi-class classification tasks. We call this method PLSR-RENT.

PLSR-RENT gives us a reduced set of features which are then used with different classifiers. The results obtained are compared with other feature selection algorithms. We observe that performance of PLSR-RENT is comparable to other feature selectors by very slight differences, though it is not better than others.

More tests need to be conducted to conclude if PLSR-RENT is the best option for extending RENT, but it is a good candidate.