A Comparison between Two Classification Algorithms by Implementing Bias-variance Decomposition

• 作 者: MeihongZhu[1]AihuaLi[2]
• 机 构: Research Center on Fictitious Economy & Data Sciences, Chinese Academy of Sciences, 100080 School of
• 出 处: 2008第四届海峡两岸应用统计学术研讨会
• 发表日期: 2008年期

We experimentally compare the performance of Multiple Criteria Linear Programming（MCLP) and Linear Discriminant Analysis(LDA)classification algorithms by implementing bias-variance decomposition. Under Domingos bias-variance decomposition framework，by using bagging ensemble，we compared their bias，variance and their variations with the size of training set on three data sets. We aimed to comparing their classification accuracy，diversity and other main characteristics. The experimental results show that，MCLP and LDA are all simple and effective classification algorithms. When training set is large enough，they present almost the same good performance.But they still behave differently in some aspects. LDA is more stable than MCLP while MCLP is more suitable for large training sets. IN their own bias-variance structures，LDA presents high bias and low variance while MCLP has the oppositional characteristics to LDA.