Maximum entropy method for multi-label classification

Details

Supervisors

Saerens, Marco

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

In machine learning, supervised classification consists in predicting a class label to real-world objects, based on data that was gathered for similar objects. In this master thesis, we focus on multi-label classification, which is a variation of supervised classification where objects can be categorized by more than one class label. In this framework, it is complicated to model thoroughly the relations between class labels, because of the exponential nature of the output. The goal of this master thesis is to develop an original method to perform multi-label classification, based on the maximum entropy principle. Thanks to this approach, we will be able to represent outputs as a whole, and therefore account for relationships between class labels. This method is implemented and tested against existing multi-label methods. Moreover, since this method will have an exponential complexity, we search for ways to reduce this complexity using heuristics.