GA-SVM Multi-Classification Algorithm Based On Dynamic Adjustment And Its Application

The problem of multi-classification is popular in practical applications and multi-classification problems have draw much attention in the field of machine learning and pattern recognition.The traditional methods of multi-classification in support vector machine such as OAO and OAA have its own shortcomings as there being a large number of classifiers,test taking too long to complete,and classification inaccuracy.Based on support vector machines with binary tree architecture for multi-class classification(BT-SVM),Scholars see the characteristics of adaptive genetic algorithm as a reference for optimization used to construct the optimal binary tree.A GA-based SVM binary-tree multi-classification method(BTA-GA)is proposed.In the BTA-GA algorithm,each node data set division directly affects the performance of the whole algorithm and there is still the problem of error accumulation.Cumulative errors tend to reduce the classification accuracy,make the classification effect become worse and there is also a problem of global optimization.The real number coding method at each node is inefficient.In this paper,we propose to construct the binary tree from the root node,and use the binary code genetic algorithm to carry out the binary tree structure of each stage.There is no need to consider the location of the binary tree in binary coding and the classification of each node and the Cross,variation become more efficient.Aiming at the cumulative effect of the error caused by the error near the root node on the subsequent node classification,a method of dynamic adjustment is proposed.This method assigns weight to each node and adjusts the weight according to the classification accuracy so that the overall classification error is reduced,And finally achieve the global optimization of the binary tree,so as to improve the classification accuracy.In order to verify the effectiveness of the algorithm,we use six UCI data sets to carry out the experiment and the five-fold cross validation to verify the algorithm stability.Compared to BT-SVM algorithm and BTA-GA algorithm,DAGA-SVM algorithm has a great improvement in global optimization and classification accuracy under the premise of the close classification time,Especially in data sets with more data categories.The feasibility and validity of the algorithm are verified.The securities investment fund is a multi-classification problem.In this paper,we select the data of stock funds in the Wind database.The processed data are applied in the traditional OAO,OAA DT-SVM and the new DAGA-SVM algorithms to compare and analyze.The experimental results show that the multi-classification algorithm proposed in this paper has practical feasibility.