The Research Of Discretization And Attribute Reduction Of Rough Sets Based On Differential Evolution Algorithm

Currently, Data Mining has been widely used in the fields of finance, manufacturing and medical care etc. But with the dramatic increase of information in the knowledge database, human beings urgently need tools that can extract potential and valuable rules from the knowledge database with large and redundant data, and interference of noise data. Since Rough sets Theory has excellent performance in these aspects, it has been applied in the Data Mining increasingly.There are a lot of continuous data in the knowledge database in practical applications, but the Rough Sets Theory based on the equivalence relation can only process the discrete data. Thus, continuous data must be discretized before using the methods of Rough Sets Theory. This paper analyzed and compared the advantages and disadvantages of unsupervised discretization methods and supervised discretization methods. Then because of the nature of optimal discretization as an NP-hard problem, it proposed a heuristic discretization method - discretization algorithm based on Differential Evolution Algorithm. The algorithm defined an individual with a real string enhancing the advantage of local search capability. On this basis, the two-value individual and a definition form of fitness function was given. The different thresholds were set on the definition of the two-value individual because of different condition attributes, and the definition of the fitness function took into account the character of the Rough Sets Theory. Finally, the experiments with Iris data set show the algorithm is effective.The attributes of the knowledge database are not equally important. You can delete irrelevant or unimportant attributes on condition that the classification of the knowledge database with or without these attributes was not changed. This paper investigated the heuristic algorithm of attribute reduction for that solving the minimum attribute reduction is an NP-hard problem. Next it proposed an attribute reduction algorithm based on Differential Evolution Algorithm and the attribute reduction effect of the algorithm was discussed in compatible and incompatible decision tables. And then a new definition form of fitness function and a new differential operation were presented for the corresponding individual of the minimum reduction. Finally, the algorithm is effective proved by two sets of experiments, and its computational complexity is the same as the attribute reduction based on Genetic Algorithm, but the convergence rate is 4 times than that based on Genetic Algorithm.