Local properties of RBF-SVM during training for incremental learning

Machine learning algorithms for large scale data are becoming more crucial in today's world. This is due to the unprecedented size of streaming data being collected by information technology. Incremental learning is considered one of the key concepts for learning from streaming data where a learned model is updated when new data becomes available in time. In this paper, we study RBF-SVM local incremental learning. The RBF-SVM decision function has been shown in the literature to have local properties which can be beneficial if they hold during learning as well. A learning machine that has local properties during learning is very desirable for incremental learning; this is because the machine will need to be updated only locally to accommodate the newly collected training data. We show via mathematical formalization and experimental verification that RBF-SVM preserves the local properties during learning. We also propose an estimate of the size of the regions in the learned model that need to be updated during the learning increments.