基于优势关系粗糙集的自主式学习模型

Self-Learning Model Based on Dominance-Based Rough Set Approach

为了增强对不一致有序信息系统的处理能力,变精度优势关系粗糙集通过引入变精度阈值增加了对不一致信息的适应性,其分类性能受变精度阈值大小的影响.然而,变精度阈值往往依赖于领域先验知识或通过反复尝试确定,极大地影响了算法的实用性.针对优势关系下如何进行信息系统知识获取这一自主控制难题,在分析了变精度优势关系粗糙集所存在问题的基础上,首先定义了优势关系信息系统中决策表的整体确定性、最大整体确定性、整体不确定性、最小整体不确定性等度量准则,进而提出了对各决策类集的最大确定性进行度量的准则和算法.在此基础上,提出了将各决策类集的最大确定性作为该决策类集的变精度阈值进行知识获取的自主式学习模型.该模型不仅避免了知识获取过程中对先验知识的依赖,也增强了对处理不一致信息系统的适应性.通过与现有算法的仿真实验对比分析,发现该自主式学习方法对处理具有较高不一致性的有序信息系统具有比较突出的优势.

In order to improve the processing ability for inconsistent preference-ordered informa- tion systems, the dominance-based rough set approach （DRSA） has been extended to variable- consistency dominance-based rough set approach （VC-DRSA）. VC-DRSA can tolerate some inconsistency through set a threshold value of consistency level and its classification performance is affected by the threshold value. The threshold value, however, is usually set according to prior domain knowledge or by a tail-and-error procedure which restricts the applications of the algorithm to a large extent. To address learning uncertain knowledge automatically driving by the monotonic information systems themselves, the inadequacies of VC-DRSA are analyzed. The integral certainty measure, maximum integral certainty measure, integral uncertainty measure and minimum integral uncertainty measure of ordered information system are defined. Furthermore, the maximum certainty of every class union is measured and a corresponding computing algorithm is proposed. Based on them, a self-learning model based on DRSA is proposed which using the max uncertainty coefficients of every class union as the consistency threshold value respectively. In this model, knowledge can be learned automatically without depending on prior domain knowledge. Besides, it strengthens the adaptability for dealing with inconsistent information systems.Through comparing with the existing algorithms, the efficiency of this self-learning method is illustrated. Especially, we found that the method has advantages for dealing with high inconsistency preference-ordered information systems.