基于免疫克隆选择的最优ECOC编码输出

An Optimization Strategy of ECOC Coding Matrix-Based on Immune Clonal Selection Algorithm

纠错输出编码(Error Correcting Output Codes,ECOC)是解决模式识别领域多类分类问题的有效工具。在寻找最优编码输出的问题上,现有方法忽略了样本类别之间的相关性,导致学习效率和分类效果低下。为构造数据感知的编码矩阵,提出基于免疫克隆选择(Immune Clonal Selection Algorithm,ICSA)的最优纠错输出编码方法,将矩阵构造的多约束NP(Non-deterministic Polynomial,NP)难问题转换为优化搜索问题.首先基于分类精度和编码长度定义亲合度函数,然后结合样本知识改进变异交叉算子,根据约束性条件对矩阵进行搜索,从而快速有效地构建最优ECOC编码.实验表明该方法能够在提升多类分类精度的同时加快算法效率,而且输出的编码矩阵更加紧凑.

Error correcting output codes(ECOC)is a powerful tool to solve multi-classification problem.The existing methods of seeking the optimal coding matrix ignore the correlation between classes,which leads to bad performance in learning speed and classification accuracy.In order to construct data-driven coding matrix,an optimization strategy of coding matrix based on immune clonal selection algorithm(ICSA)is presented.The strategy reduces the multiple constraints non-deterministic polynomial problem(NP)of finding the optimal coding matrix to a finite heuristic search problem.Firstly,the affinity function based on accuracy and coding length is defined.Then,the mutation,crossover and selection operator are modified respectively.Meanwhile,the validity constraints are combined to execute the quick search.The experiment results based on UCI and traffic data prove that the proposed strategy can enhance the classification performance and accelerated the speed.The output coding matrix is more compact as well.