Research On Classification Rule Mining Based On Genetic Algorithms

Data Mining is a new research realm rising in recent years, and synthesizes many fields of science such as the database technique, artificial intelligence and statistics etc. Its main job is to discover valuable information and knowledge hidden behind data from numerous data in order for decision support. Mining classification rules is a procedure to construct a classifier through studying training dataset, and is a very important part of Data Mining and Knowledge Discovery. In essence, its goal is to discover classification rules which are highly accurate, comprehensible and interesting. Genetic Algorithm (GA) is an overall random searching method based on the theory of evolution and molecular genetics.This thesis has done some research on classification based on Genetic Algorithm, proposes a new operator called similarity-based crossover-mutation, improves a genetic algorithm based on the competition between populations, and applies the new operator and the improved algorithm to mine the classification rules.In this thesis, we look back the history background of data mining at first, and sum up the basic concepts, process, characteristics, classification and the mission modes of data mining in detail, emphasizing to discuss the steps, techniques and methods for classification rule mining. Then, this thesis introduces the biologic origin, the characteristics and the theories foundations of GA, sums up the process of the Simple Genetic Algorithm (SGA), three operators and four key problems. Also, it analyzes and discusses the reason and resolution of the premature phenomena when SGA is applied in mining classification rules. Furthermore, in order to overcome the premature phenomena, based on the SGA, this thesis introduces the idea of "similarity degree" and "beneficial crossover", and puts forwards a new operator called similarity-based crossover-mutation by combining crossover with mutation together, and applies the new operator into mining classification rules on Breast cancer data. Finally, this thesis improves a genetic algorithm based on the competition between populations, and tests the improved algorithm by adult data.