In this paper we make an extensive study of different combinations of ensemble techniques for improving the performance of adaboost considering the following strategies: reducing the correlation problem among the features, reducing the effect of the outliers in adaboost training, and proposing an efficient way for selecting/weighing the weak learners. First, we show that random subspace works well coupled with several adaboost techniques. Second, we show that an ensemble based on training perturbation using editing methods (to reduce the importance of the outliers) further improves performance. We examine the robustness of the new approach by applying it to a number of benchmark datasets representing a range of different problems. We find that compared with other state-of-the-art classifiers our proposed method performs consistently well across all the tested datasets. One useful finding is that this approach obtains a performance similar to support vector machine (SVM), using the well-known LibSVM implementation, even when both kernel selection and various parameters of SVM are carefully tuned for each dataset. The main drawback of the proposed approach is the computation time, which is high as a result of combining the different ensemble techniques. We have also tested the fusion between our selected committee of adaboost with SVM (again using the widely tested LibSVM tool) where the parameters of SVM are tuned for each dataset. We find that the fusion between SVM and a committee of adaboost (i.e., a heterogeneous ensemble) statistically outperforms the most used SVM tool with parameters tuned for each dataset. The MATLAB code of our best approach is available at bias.csr.unibo.itanni/ADA.rar .
展开▼
