A Boosting Method with Gaussian Mixtures as Base Learners in a Low-Dimension Space

In this paper we propose a classification method in the context of Boosting called Transformed Space Boosting (TSB). Our aim is to develop the idea of using a combination of Gaussian Mixture Models and transformation matrices to design ´non-weak´ base learners in Boosting strategies. The use of transformation matrices makes it possible to do a linear dimensionality reduction from an original space to a transformed one. This leads to a two-steps method where in the first one a single-component mixture is trained in the original space. In the second step, based on the single Gaussian previously trained, we apply the concept of average divergence measure to estimate the transformation matrix. The final classifier achieves an improvement in performance compared to other methods also based on dimensionality reduction. This is clearly seen from the experiments we present which strength the validity of our method and show promising classification scores.