A Rule Extraction Method for Hybrid Data 

Yan LI1, Tingting WU2 

Key Lab. Of Machine Learning and Computational Intelligence, College of Mathematics and Information Science, Hebei University, Baoding 071002, China

Abstract: In real world applications, the collected data often consists different types of attributes such as symbolic and ordinal attributes, which can be considered as hybrid data. In rough sets, the equiva-lence/dominance relations are used to represent and approximate data classification for symbolic data and ｏｒ-dinal data respectively. In order to extract decision rules in hybrid data, some scholars have proposed the me-chanism to generate monotonic rules based on dominance relations. The original symbolic symbol values are also treated as preference-ordered attributes. Although this approach can generate rules with higher coverage, the efficiency is not satisfactory because all symbolic attributes need to be sorted two times as gain-type and cost-type. In this paper, we define dominance-equivalence relations on conditional attributes of hybrid data which introduces equivalence relations on symbolic attributes and dominance relations on ordinal attributes respectively. Thus, this method could preserve the original meaning of the data, and the rule generation could also be accelerated. Based on dominating and dominated classes, the upper and lower approximations and the decision rules can be computed and the matching mechanism of the rules is also established. Eleven UCI data sets are selected in the conducted experiments. The results show that the proposed method not only extracts more rules than the monotonic rules but also can reduce the running time obviously while the classification precision is also improved slightly. 

Keywords: Rough set; Dominance relation; Equivalence relation; Monotonic Rules