Attribute reduction via local conditional entropy
- PDF / 2,780,527 Bytes
- 16 Pages / 595.276 x 790.866 pts Page_size
- 90 Downloads / 221 Views
ORIGINAL ARTICLE
Attribute reduction via local conditional entropy Yibo Wang1 · Xiangjian Chen1 · Kai Dong1
© Springer-Verlag GmbH Germany, part of Springer Nature 2019
Abstract In rough set theory, the concept of conditional entropy has been widely accepted for studying the problem of attribute reduction. If a searching strategy is given to find reduct, then the value of conditional entropy can also be used to evaluate the significance of the candidate attribute in the process of searching. However, traditional conditional entropy is used to characterize the relationship between conditional attributes and decision attribute in terms of all samples in data, it does not take such relationship with specific samples (samples with same label) into account. To fill such a gap, a new form of conditional entropy which is termed as Local Conditional Entropy is proposed. Furthermore, based on some important properties about local conditional entropy studied, local conditional entropy based attribute reduction is defined. Immediately, an ensemble strategy is introduced into the heuristic process for searching reduct, which is realized by the significance based on local conditional entropy. Finally, the experimental results over 18 UCI data sets show us that local conditional entropy based attribute reduction is superior to traditional conditional entropy based attribute reduction, the former may provide us attributes with higher classification accuracies. In addition, if local conditional entropy is regarded as the measurement in online feature selection, then it not only offers us better classification performance, but also requires lesser elapsed time to complete the process of online feature selection. This study suggests new trends for considering attribute reduction and provides guidelines for designing new measurements and related algorithms. Keywords Attribute reduction · Conditional entropy · Local conditional entropy · Neighborhood rough set · Rough set
1 Introduction Rough set theory [1], was firstly proposed by Pawlak. Classical rough set model is useful in characterizing the uncertainty in data with categorical values by the construction of lower and upper approximations. Presently, rough set and its various generalizations [2–8] have been demonstrated to be effective in solving practical problems related to Machine Learning, Pattern Recognition, Social Network, Knowledge Discovery [9] and so on.
* Kai Dong [email protected] Yibo Wang [email protected] Xiangjian Chen [email protected] 1
Though this form of approximation is the basic foundation of rough set theory, it should be noticed that attribute reduction has contributed much to the development of rough set theory. Attribute reduction, also called the rough set based feature selection [10–15], can be easily distinguished with other feature selection techniques because the former has clear explanations. For example, given a measurement related to rough set, attribute reduction can be described as a minimal subset of the raw attributes which meets the
Data Loading...
