FAST SVM TRAINING USING APPROXIMATE EXTREME POINTS By MANU NANDAN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY

of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy FAST SVM TRAINING USING APPROXIMATE EXTREME POINTS By Manu Nandan December 2013 Chair: Pramod Khargonekar Cochair: Sachin Talathi Major: Computer Engineering Support vectors machines (SVMs) are widely applied machine learning algorithms that have many desirable characteristics. However, with the ever increasing size of datasets, it has become increasingly difficult to use SVMs in modern applications. In this dissertation we address two disadvantages of SVMs. Firstly, non-linear kernel SVM solvers need excessive training times with large datasets. Secondly, linear SVM solvers need to be parallelized to train on web-scale datasets that are large and high dimensional. State-of-the-art linear SVM solvers, though fast, are difficult to parallelize. We propose a modification, called the approximate extreme points support vector machine (AESVM), that is aimed at overcoming these disadvantages. Our approach relies on conducting the SVM optimization over a carefully selected subset, called the representative set, of the training dataset. We present analytical results that indicate the similarity of AESVM and SVM solutions. Linear or log-linear time algorithms based on convex hulls and extreme points are used to compute the representative sets. We also propose an algorithm to post-process the solution of SVMs to enable fast classification. A variant of the algorithm is easy to parallelize and is designed to compute efficiently on frameworks such as MapReduce. Extensive computational experiments on thirteen datasets compared our algorithms to other modern SVM solvers. We compared our non-linear AESVM solver to LIBSVM [15], CVM [76] , BVM [75], LASVM [7], SVM [36], and the random features method [61]. Our AESVM