LARGE-SCALE CLASSIFICATION BASED ON SUPPORT VECTOR MACHINE

This thesis proposes the fast support vector classifier, an efficient implementation of the radial basis function support vector machine (SVM) for large- scale classification problems. It achieves performance near the state-of-the-art, being much faster than existing approaches over datasets up to 31 million patterns, 30,000 inputs and 131 classes. It also adjusts the memory requirements, allowing to be executed on datasets of almost arbitrary size. The thesis also proposes the ideal kernel tuning, an efficient tuning method for the Gaussian kernel spread of the SVM, that is the fastest compared to other five methods in the literature, with performance very near to the best and reduced memory requirements.

Palabras clave: Support vector machine, hyper-parameter tuning, high dimensionality ., Machine learning, classification, large-scale data classification multi-class