Title: Consistency and robustness properties of predictors based on locally learnt support vector machines
Authors: Florian Dumpert - University of Bayreuth (Germany) [presenting]
Abstract: Among different machine learning methods, support vector machines (SVMs) play an important role in many fields of science nowadays. A lot of research about statistical and computational properties of support vector machines and related kernel methods has been done during the last two decades up to now. On the one hand, from a statistical point of view, one is interested in consistency and robustness of the method. On the other hand, from a computational point of view, one is interested in a method that can deal with many observations and many features. As SVMs need a lot of computing power and storage capacity, different ways to handle big data sets were proposed. One of them, which is called regionalization, divides the space of the declaring variables into possibly overlapping regions in a data driven way and defines the output predicting function by composing locally learnt support vector machines. It is possible to show that a predictor learnt in this way conserves consistency and robustness results under assumptions that can be checked by the user of this method. We will have a closer look at the consistency and robustness properties of predictors based on locally learnt support vector machines.