An empirical study of Conserved Self Pattern Recognition Algorithm: Comparing to other one-class classifiers and evaluating with random number generators

Early work has demonstrated that conserve self pattern recognition algorithm (CSPRA) produces promising performance in the field of anomaly detection. This paper further extends the applications of CSPRA to Fisher´s Iris data, Indian Telugu data and Wisconsin breast cancer data. A formal description of the differences between the two detection strategies (classical CSPRA and selective CSPRA) is given and the results show that selective CSPRA performs better for the tested data. The comparative study of CSPRA to other one-class classifiers (NSA, V-detector and One-class SVM) shows that the performance of the CSPRA is obviously better. This paper also investigates the influence of various random number generators on the performance of the CSPRA and NSA. Our experiments indicates that non-uniform random number generators tend to produce worse performance than uniform random number generators and quasi-random number generator has the potential to enhance the system performance compared to other uniform random number generators.