Incremental Real Time Support Vector Machines

Abstract

This paper investigates the problem of handling large data stream and adding new attributes over time. We propose a new approach that employs the dynamic learning when classifying dynamic datasets. Our proposal consists of the incremental real time support vector machines (I-RTSVM) which is an improved version of the support vector machines (SVM) and LASVM. On one hand, the I-RTSVM handles large databases and uses the model produced by the LASVM to train data. It updates this model to be appropriate to new observations in test phase without re-training. On the other hand, the I-RTSVM presents a dynamic approach that adds attributes over time. It uses the final model of classification and updates it with new attributes without re-training from the beginning. Experiments are illustrated using real-world UCI databases and by applying different evaluation criteria. Results of comparison between the I-RTSVM and other approaches mainly the SVM and LASVM shows the efficiency of our proposal.