Resource-limited Parallel SVM

Support Vector Machines (SVMs) need large memory requirement and computa-tion time when dealing with large datasets. Present work has not taken the memorylimit into consideration and assumed the problem was solved on resource-unlimitedenvironment. However, the problem usually need to be solved on a resource-limitedenvironment. Existing work are not resource-limited. To solve this problem, we pro-pose resource-limited parallel SVM. Our main work includes:1. We design and implement an efficient parallel algorithm, RF-CCASVM, forSVMS, which uses Random Fourier features and consensus centre adjustmentstrategy.2. We propose a resource-limited scheme for parallel SVMs. We explicitly mapdata into low-dimensional features space via random Fourier mapping, and traina linear SVM in the space.3. We derive an error bound for the approximate algorithm, and analyze the prop-erty of the key parameter for the resource-limited scheme.