A Study Of Classification Based On Particle Swarm Optimization

Classification is a task aiming to find a concept description extracted from the training data, setting up a classifier, and predicting unlabeled instances. As an important branch of data mining, it has been widely applied into many areas, such as business, medicine, military affairs, etc. The common classification methods have decision tree, Bayesian network, neural network, rough set, fuzzy set, support vector machine and so on. As a special variant of computational intelligence family, particle swarm optimization (PSO) has been widely applied into many areas because of its simple concept and easy implementation. In this article, PSO is applied to extract classification rule aiming at numerical attributes.Because the performance of PSO is sensitive to the control parameters, this article proposes a new social parameter adjustment strategy. It adds social weight upon velocity information for an increased iteration, and is applied to extract classification rule, simulation results show the proposed social strategy is superior to standard PSO. Secondly, the fitness function selection plays an important role in classification. The current literature for fitness function maintains some shortages. To overcome these disadvantages, this article introduces a new fitness function, and successfully applied into UCI datasets, simulation results show it is better than other fitness functions mentioned before.