关键词: 支持向量数据描述, 改进磷虾群算法, 参数优化, 精英选择和保留策略

Abstract: Support Vector Data Description（SVDD） is a classification algorithm for constructing one-class data description. Penalty parameter[C]and kernel parameter[σ]are two key points to affect the classification performance of SVDD and it has always been a difficulty to select reasonably. Focused on this problem, a parameter optimization algorithm of SVDD based on improved krill algorithm is proposed. Firstly, this paper analyzes the[C]and[σ]parameters’ influence on learning models of SVDD through simulation experiment; Then, this paper introduces the krill herd algorithm and analyzes its merits and demerits, defines the disturbance factor in random diffusion behaviors to enhance the global exploratory; Furthermore, a new elitist election and reserving strategy is introduced into the iterative process to improve the accuracy of convergence; Finally, this paper introduces the improved krill herd algorithm into the parameters optimization process of SVDD and establishes the IKH-SVDD parameters optimization model. Simulation with the UCI benchmark datasets indicates that KH-SVDD has better classification accuracy than the current parameter optimization algorithm.

Key words: support vector data description, improved krill herd algorithm, parameters optimization, elitist election and reserving strategy