| Classification is one of the key contents in the field of data mining.The main goal of classification mining is to train a model with strong generalization ability in the sample set of known categories,so as to have an accurate prediction of new data.At present,although researchers have done a lot of research in order to improve the accuracy of prediction and achieved a series of results,there are still many problems worthy of in-depth study and discussion.In this paper,the characteristics of classification data selection,comparative analysis of classification algorithm and the ensemble of classification algorithm are studied,and the experimental results are verified.Firstly,the paper discusses the feature selection of classification data.An improved mRMR algorithm is proposed to solve the problem that the mRMR feature selection algorithm has fixed feature evaluation criteria for all data sets,while ignoring the different redundancy and correlation of different data sets.By adding the weight factor,the algorithm can adjust the ratio of the maximum correlation and the minimum redundancy in the evaluation standards of different data sets,so as to better fit the characteristics of different data sets.The experimental results of UCI data set are used to compare the effectiveness of the method.Then,in view of the large number of classification algorithms,people don't know how to determine the algorithm in the face of practical problems.For nine typical classification algorithms,starting with the classification performance influencing factor of the number of data sets,the performance difference between the binary classification problems and the multiclass classification problems is analyzed by using the experimental method.In the comparative experiment,aiming at the two problems,17 data sets are selected from UCI data set respectively.On the basis of data preprocessing and algorithm parameter optimization,the comprehensive performance of 9 algorithms on the two data sets is tested respectively.The experimental results are analyzed and evaluated from four aspects of classification precision,classification efficiency,scalability and robustness Price.Finally,in order to solve the limitation of single classification algorithm in improving prediction accuracy,this paper uses ensemble learning technology to improve its generalization ability,and constructs an ensemble classification algorithm based on improved particle swarm optimization algorithm.In this method,aiming at the disadvantage that the particle swarm optimization algorithm with linear decreasing inertia factor can not balance the global search ability and local search ability,and that the particle search process is not consistent with the actual nonlinear change characteristics,an improved particle swarm optimization algorithm with nonlinear S-type change of inertia factor is proposed,which can balance the search ability by highlighting the global search ability in the early stage and the local search ability in the later stage The cable capacity accelerates the convergence speed.On this basis,the feasibility and effectiveness of the proposed method are verified by simulation experiments. |
