Abstract: Without taking the spread of negative class samples into account, the objective of single spherical classifier (RSS) is only to maximize the separation ratio. According to the Fisher discriminant analysis, this paper introduced relafive margin into RSS to enhance the cohesion of negative class samples and improve the discriminant accuracy by the upper bound constraint in the feature space. Because the upper bound is unpredictable, a maximum relative separation ratio single spherical classifier with an adaptive upper bound (ARRSS) was built to avoid no solution and its parameters were researched afterwards. Experiments show the proposed method achieves better generalization performance compared with RSS.

Key words: RSS, Fisher discriminant analysis, Relative margin, Upper bound constraint, Adaptive upper bound