Discretization Based On Rough Sets Theory

Discretization of continuous attributes which is an important part of datapreprocessing, is directly related to the effect of data mining or machine learning. Theprocess of discretization based on rough sets is generally divided into three steps: theextraction of candidate cuts, the selection of result cuts from candidate cuts, and theapplication of result cuts for discretization. The first step for the extraction of candidatecuts is basis for the solvelution of the discretization problem. Under the premise ofensuring the compatibility of the decision table, to make the candidate cuts has thesmallest possible base, not only can reduce the calculation time of the result cuts, butalso can reduce the calculation time and space expenses of the discretization process, ithas great significance for the successor of the entire discretization algorithm. The mainwork of this paper is as follows:1) The detailed analysis of the three steps of the process of discretization based onrough sets: the extraction of candidate cuts, the selection of result cuts fromcandidate cuts, and the application of result cuts for discretization.Among them,the first two steps is the key of the discretization algorithm, but the existingdiscrete algorithm study focused onthe selection of result cuts from candidatecuts, while ignoring the extraction ofcandidatecuts.The main research contentsin this paper is the extraction of candidate cuts which were analyzed in detail.2) The process of candidate cuts sets extraction is analyzed, by the analysis, a newalgorithm for candidate cuts is proposed based on shadowed sets.According todistribution of the instance, the algorithm sort the instance, and using shadowedsets for calculating every class’s upper and lower approximation, at last thecandidate cuts is extracted.Several UCI data sets is applied to test theperformance of the algorithm and the experiment result is compared with otheralgorithm for candidate cuts. The experiment result shows that the algorithmcan effectively reduce the number of candidate cuts in data sets, acceleratethe speed of discretization and the rate of recognition.3) The discretization algorithm based on shadowed sets is applied in radar emitterrecognition. Throughanalyzing the relationship of discretization and attributereduction in rough sets，it is acquaintanced that discretization and attributereduction are all reduction on equivalence in essence, so it is necessary toimprove the radar emitter recognition model structurebased on rough sets. Atlast, using the improved model structure to recognize the radar emitter featuredatabase, the results of experiment prove that the feature selection model can efficiently extract the useful information from data, simplify the structure ofdata, consequently it simplifiys the structure of Neural Network, shortens thetraining time of classification, enhances the generalization capability ofclassifier.