Research On SVM Decision-tree Multi-classification Strategy Based On Electromagnetism-like Mechanism

In recent years, Multi-Classification problem is one of the hot topics in the study ofsupport vector machine. Among some proposed algorithms so far,1-a-r algorithm and1-a-1algorithm are poor in training and classification efficiency when the number ofcategories is large, and both exist "unrecognized area". DAG-SVM algorithm has twodisadvantages, one is slow training speed for many classes of classification problems,and another is the selection of root node will directly affect the final classification result.DT-SVM algorithm has a higher training and classification speed with no "unrecognizedarea" existing, but due to its fixed tree structure and the randomly selection of eachdecision node, it can easily produce "error accumulation", thus the classificationperformance is not stable, and the result is hard to achieve optima. GADT-SVMalgorithm uses genetic algorithms to optimize the structure of the tree, hence, aself-adaptive structure tree is obtained, but its classification accuracy is still not highenough.In order to further improve the overall performance of SVM Multi-Classificationstrategy, this paper proposes a SVM Decision-Tree Multi-Classification algorithm basedon Electromagnetism-like Mechanism(EM) which uses EM to guarantee the two classesis optimal at every decision node, thereby an optimal or near-optimal decision-tree canbe created automatically. First, an improved Electromagnetism-like mechanismalgorithm is proposed based on the characteristics of the classification problem; byanalyzing Multi-Classification problem and all kinds of coding schema, Real NumberCoding strategy is used to encode the particle; according to the force of the particles, anew moving method of particles which fits the characteristics of theMulti-Classification problems is proposed here; an objective function is designed basedon the maximal margin between two classes. Then an EM-based decision tree algorithmis designed to deal with SVM Multi-Classification problem. Finally, the new algorithmis analyzed theoretically and simulated in this paper.Experimental results show that the improved algorithm can create optimal or near-optimal decision tree and be applied to the Multi-Classification problems successfully.It has a better overall performance than1-a-1,1-a-r, DAG-SVM, DT-SVM andGADT-SVM algorithm.In further research, the EM algorithm would be further improved on discrete optimization, and the improved algorithms would be used to solve the optimizationproblems in practical projects.