A novel local success weighted ensemble classifier

Ensemble methods aggregate the decisions of diverse component classifiers to achieve superior classification performances. Most of the previous ensemble frameworks have used fixed weights to determine the influence of each of the component classifiers on the ensemble decision. However, in practice base classifiers usually have expertise in local regions of the feature space. This paper presents a novel framework for instance dependent weighing of base classifiers instead of fixed weights. These classifier weighing functions are linked to the ability of the respective classifiers to correctly predict the labels. For a particular test pattern, base classifiers with higher likelihood of predicting the correct label have higher weights. Thus, weighing functions curb (by having lower weights) the redundant false decisions responsible for mis-classification. Regression models are used to learn the weighing functions. Experimental results on well-known benchmark datasets support the effectiveness of the proposed approach. However, the use of regression models to determine the weights makes the operations slightly computationally expensive.