Reducing the number of support vectors to allay inefficiency of large-scale models in computational biology

Support vector machines (SVMs) are powerful machine learning techniques that have been applied to numerous modelling problems, including classification and prediction models for biological sequence data. SVMs can suffer from slow classification times due to large numbers of support vectors and computationally expensive kernel functions. This paper shows that some SVMs can have support vectors removed to improve classification time, with little to no loss in generalisation accuracy. Greedy and stochastic search methods for finding the support vectors to remove are tested on sets of biological sequence data. This paper shows the results of reducing SVMs trained on sequence data sets, where accuracy was retained after more than half the support vectors were removed