MCCI_ 利用倾向性函数挖掘紧凑对比不等式.docVIP

  MCCI: Mining Compact Contrast Inequality with Preferred Functions# 5 10 15 20 25 30 35 DUAN Lei, TANG Changjie, ZUO Jie, ZHANG Tianqing* (School of Computer Science, Sichuan University, Chengdu 610065) Abstract: Contrast inequality is a new analysis tool to describe the contrast relationship between classes of data in the form of inequality. A Gene Expression Programming (GEP) based method has been proposed for contrast inequality mining in previous work. However, the method does not consider the comprehensibility and clarity of the discovered contrast inequality. The main contributions of this paper include: (1) introducing a compactness measure for contrast inequality, (2) designing a method, named MCCI, to discover compact contrast inequalities with preferred functions, (3) discussing some strategies used in MCCI to keep the compactness measure efficient, and (4) giving an extensive performance study on several data sets from UCI Machine Learning Repository. The experimental results show that MCCI is effective for mining compact contrast inequalities with high discriminative power. And domain knowledge can be introduced in MCCI to evolve inequalities containing preferred functions. Key words: Data Mining; Contrast Mining; Contrast Inequality; Compactness Measure 0 Introduction The significant characteristics and contrasts between different classes of data are interesting for scientists and useful for other data mining task, such as classification. A well known method to describe this kind of contrast relationship is Emerging Pattern (EP), which is proposed by Dong and Li in [1]. Since the concept of EP is proposed, many works on EP discovery, EP extension and application have been deeply researched [2-10]. Contrast Inequality is a new kind of functional contrast proposed in our previous work [11]. Different from EPs, which are conjunctions of items that have high support in one class and low support in other class, a contrast inequality is composed of one