Research On Evolutionary Computation Based Complex Classification Algorithms

Recently, the amount of data stored in computers has been increasing tremendously as the result of the development of information technology and applications. Data mining provides techniques that extract potential useful information from large databases. As a new intelligent and optimization technique, the evolutionary computation has been widely applied in data mining tasks. In this thesis, the important problems in complex classification data mining tasks are studied by adopting the latest evolutionary computation methods. The main contents and contributions of this thesis includes:(1) New methods and hot research subjects are introduced, especially in classification data mining domain. The basic theory and recent developments of evolutionary computation are also introduced.(2) A hybrid evolutionary algorithm for designing fuzzy classifier is proposed. The algorithm uses a novel niching technique to maintain the diversity of population, which obtains all rules of fuzzy classifiers in a single run of the evolution algorithm. A local search method is designed in the algorithm, which is able to improve the quality of learned rules. Experiments show that the local search method can effectively improve the fitness of the population, and the algorithm can produce fuzzy classifiers with low complexity and high prediction accuracy.(3) A discriminative classification algorithm based on genetic programming is proposed. Classifiers based on genetic programming is good at handling two-class problems, but the ability of solving multi-class problem is limited. We present a new classification model—discriminative classification model. The new model, optimized by minimizing the square error, can generate highly accurate classifier. In addition, an ensemble method of the new model is proposed, which can further improve the classifier's performance. Experiments show that the discriminative classification model predicts the unlabeled instances with a bigger accuracy, while the ensemble method can enhance the performance of the new model.(4) A feature selection algorithm based on multiobjective evolutionary algorithm is developed. There are two important aspects of the feature subset quality, one is feature relevance and the other is feature redundancy. A new feature redundancy measurement is proposed which evaluates the redundancy of feature subset by class-dependent mutual information; then a multiobjective evolutionary algorithm is employed to maximize the feature relevance and minimize the feature redundancy simultaneously. Experiments show the new redundancy measure can improve the performance of greedy feature selection algorithm, and the feature subsets generated by multiobjective evolutionary algorithm are better than those produced by greedy algorithms.