Multi-class Classification Algorithm Based On Logistic Regression And Support Vector Machine

Regression and classification is a commonly used method for quantitative analysisof data. Usually, it is necessary to build a regression model to analyze the relationshipbetween the explanatory variables and the response variable in the analysis of the data.However, the quantitative regression analysis method is not useful when the responsevariable is an attribute indicator, then it needs building a Logistic regression model toanalyze the relationship between these variables. It is suitable to solve the classificationproblems about the Logistic regression model. Nowadays, there are many classificationalgorithms in classification field. Two of them are Logistic regression and supportvector machine (SVM). It uses a probability to be classified in Logistic regressionmodel, and its advantages are adaptable, good robustness and good explanation. SVM’smain advantage is higher prediction accuracy and it has good performance in solvingthese problems of small sample size, nonlinear and high-dimensional vector. This is thefocus of the study that classification by multi-class classification algorithms in thepresent age.Traditional multinomial Logistic regression (MLR) model is based on the outputprobability as a classification standard, which may have a greater miscarriage whenthere are a few probabilities near1/K (K is the number of classification).To solve thisproblem, multi-class classification SVM (MSVM) model is pulled in MLR model, thenan integrated multi-class classification algorithm (MLR-MSVM) by MLR and MSVMmodel is proposed in this paper. The output of MSVM is the basis concept of the outputprobability of MLR, then it reduces the miscarriage risk and improves the correct andefficiency in classification and discrimination. UCI machine learning repository data isused in experimental analysis. Compared with the MLR, MSVM and MLR-MSVMalgorithm, the experimental results show the integration algorithm has best robustnessand classification results, and it is an effective algorithm.