Towards robust SVM training from weakly labeled large data sets

Learning from large data sets that contain samples of unknown or incorrect labels becomes increasingly important. Such problems are inherent to many big data scenarios, hence there is a need for developing robust generic approaches to learning from difficult data. In this paper, we propose a new memetic algorithm that evolves samples and labels to select a training set for support vector machines from large, weakly-labeled sets. Our extensive experimental study confirmed that the new method presents high robustness against weakly-labeled data and outperforms other state-of-the-art algorithms.