Research On Evolutionary Mining Of Classification Ruleset

Data mining is the process of automatic extraction of novel, useful and understandable patterns in large databases. Classification rules mining is one of the most common forms of knowledge discovery. It is a method for discovering a set of"if/then"rules that can be used for classification or estimation. Many studies have shown that rule-based classification algorithms perform very well in classifying both categorical databases and sparse high-dimensional databases. In this dissertation, rule mining methods based on evoloution algorithms are investigated. The research work is as follows:Various conventional classification methods for data mining task are reviewed, including rule-based classification methods. We also analyze rule mining methods based on evolution algorithm, and then present the research plan of this dissertation. A multi-population coevolutionary framework is employed to evolve multiple rule species in parallel. Each population aims to find a rule which can accurately classify part of the training instances. When the quality of an individual is measured, collaborators from all the other populations have to be selected with the aim of forming complete ruleset. Therefore individuals in different populations are encouraged to cover different part of the training instances so as to form a better ruleset. The algorithm also allows the number and roles of the populations to be adapted during evolution process, so more accurate and general rules can be generated.A new coevolutionary genetic algorithm is proposed to mine multiple rules in one population. The algorithm assigns fitness to individuals according to their coverage of training instances, which makes the fitness of one individual dependent on other individuals'performance. Therefore cooperation between rules is promoted, and more accurate rule set can be generated with less computation time.