基于Reduct的“规则+例外”学习

'Rule+Exception' Learning Based on Reduct

一般地说,机器学习关注的是“规则”,并将规则不能覆盖的“例外”考虑为噪声.然而,大量的应用不仅需要刻画满足大多数观察的规则,同时需要显现可解释地表示例外.在情报分析与安全预警这类应用中,例外可能是更为重要的知识.对此作者描述了一类限制在结构化符号数据集合上的基于Reduct的“规则+例外”学习的理论框架,并给出了解决这个框架各个组成部分中所存在的问题的一个方案.

Machine learning normally focuses on “rules” and treats “exceptions” not covered by the rules as noise. In many applications, it is necessary to have not only rules describing the observations but also explicit and explainable representation of exceptions. Exceptions may be an important type of knowledge in these applications, such as intelligence analysis and security warning. Based on the notion of Reducts, a theoretical framework for learning “rule＋exception” knowledge is presented within the context of symbolic learning. The basic components and the main issues of the framework are discussed.