A Study Of Evolving Classification Rule Based On PSO Algorithm

The classification is a kind of technology of data mining and has already obtainedapplication in the trade. The common classification algorithm has decision tree, statisticalmethod, machine learning, neural network and so on. Since genetic classification systemwas implemented, many evolution algorithms were applied to classifier because they showtheir advantage in solving complex problems.Particle Swarm Optimization (PSO) algorithm has its own special characters as one ofevolution algorithms. Based on analysis of PSO model and classification model, PSOclassification system was presented and showed its efficiency through the experiment resultsin this paper. Firstly, coding, fitness function and entire structure of PSO in classificationsystem was defined. PSO of Single population was applied to implement classification system.Secondly, multi-category problem was taken as multi-two-class problem, a new approach ofusing multi-particle swarm to evolve multi-classification rule was proposed in order toimprove algorithm performance. In this way, a population expressed a category andmulti-category was indicated by multi-particle-swarm. Then, hybrid code consisted of realand binary number was designed to represent 'no care ' attribute .As a result, it reducedclassification time. Finally, on the basis of analysis of PSO and Genetic Programming (GP),PSO and GP was combined to evolve rule which GP was used to evolve rule structure andPSO was applied to evolve rule value. Because PSO tends to run into local optimum in latter period, PSO was improved topreferable solve classification. PSO with adaptive and stochastic inertia weight waspresented and mutation operator was introduced. They all obtain better experiment results.