Using Locality-Sensitive Hashing for SVM Classification of Large Data Sets

We propose a novel method using Locality-Sensitive Hashing (LSH) for solving the optimization problem that arises in training stage of support vector machines large data sets, possibly high dimensions. LSH was introduced as an efficient way to look neighbors dimensional spaces. Random projections-based functions create bins so when great probability points belonging same bin are close, far will not be bin. Based on these bins, it is necessary consider whole original set but representatives each one them, thus reducing effective size set. A key our proposal we work with feature space and use only projections search closeness this space. Moreover, instead choosing projection directions at random, sample small subset solve associated SVM problem. Projections direction allows more precise many cases approximation solution found fraction running time degradation classification error. present two algorithms, theoretical support, numerical experiments showing their performances real life problems taken from LIBSVM base.