Training Support Vector Machine can become very challenging in large scale problems. Training several lower complexity SVMs on local subsets of the training set can significantly reduce the training complexity and also improve the classification performances. In order to obtain efficient multiple classifiers systems, classifiers need to be both diverse and individually accurate. In this paper we propose an algorithm for training ensembles of SVMs by taking into account diversity between each parallel classifier. For this, we use an information theoretic criterion that expresses a trade-off between individual accuracy and diversity. The parallel SVMs are trained jointly using an adaptation of the Kernel-Adatron algorithm for learning online multiple SVMs. The results are compared to standard multiple SVMs techniques on reference large scale datasets.
