A review and comparison of multi-target classification techniques

Miran Levar (2012) A review and comparison of multi-target classification techniques. EngD thesis.

Preview

PDF Download (776Kb)

Abstract

Does simultaneous classification of multiple target variables perform better than building a classifier for each of the target variables independently? To answer this question we implemented a set of classification techniques for multi-target classification and integrated them into Orange Multitarget, an add-on for Orange, an open-source machine learning framework. Performance of both multi-target (clustering trees, neural networks, PLS) and single-target techniques (e.g. random forests) was tested on multiple datasets, which included datasets with binary class variables and datasets with multinomial class variables. The results do not show an advantage for either of the techniques. We have also observed that increased correlation between class variables does not increase the performance of multi-target techniques when compared to single-target techniques.

Actions (login required)