Parameter estimation of support vector machine with radial basis function kernel using grid search with leave-p-out cross validation for classification of motion patterns of subviral particles

Abstract

The classification of subviral particle motion in fluorescence microscopy video sequences is relevant to drug development. This work introduces a method for estimating parameters for support vector machines (SVMs) with radial basis function (RBF) kernels using grid search with leave-pout cross-validation for classification of subviral particle motion patterns. RBF-SVM was trained and tested with a large number of combinations of expert-evaluated training and test data sets for different RBF-SVM parameters using grid search. For each subtest, the mean and standard deviation of the accuracy of the RBF-SVM were calculated. The RBF-SVM parameters are selected according to the optimal accuracy. For the optimal parameters, the accuracy is 89% +- 13% for N = 100. Using the introduced computer intensive machine learning parameter adjustment method, an RBF-SVM has been successfully trained to classify the motion patterns of subviral particles into chaotic, moderate and linear movements.

Keywords

• grid search
• leave-p-out cross-validation
• parameter estimation
• principal component analysis
• subviral particles
• support vector machine