The Research Of Classification Algorithms Based On SVM And Feature Combination

In the modern society,data information grows in the form of explosion.The sample size and dimension of data are increasing continuously.The emergence of data mining technology effectively solves the problem of how to find valuable information from massive data.Classification is one of the most important techniques in data mining and is widely used in many fields.How to make use of feature information effectively and improve the effect of classification algorithm is a hotspot in the research of classification technology.It is proved that feature combination can effectively improve the classification performance of the classification algorithm.TSP(Top Scoring Pair)algorithm combines features in the form of feature pairs,and selects the optimal feature pairs for classification.It has the advantages of simple and efficient.k-TSP is an extension of TSP algorithm.Different from TSP,k-TSP uses k>0 feature pairs for classification.From the perspective of feature combination,TSP uses a fixed combinatorial form for any two features.It is a special form of linear combination of features.The constraint of combination form limits the classification ability of the feature pairs in TSP algorithm.In this paper,LC-TSP algorithm is proposed which adopts SVM(Support Vector Machine)to construct the linear combination of two features,which is used to replace the fixed combination form in TSP algorithm.LC-TSP is extended to LC-k-TSP algorithm with using multiple combination features for classification.Experiments on public datasets show that LC-TSP and LC-k-TSP have better classification performance than TSP and k-TSP.The combinatorial feature of correlation is constructed based on the Pearson correlation coefficient of a pair of features.It has been proved that it can represent the change of correlation between features,which is helpful to discover better features for classification.However,the construction of the combinatorial feature of correlation will add too many features to the feature space,which may cause the curse of dimensionality.This paper presents the correlation kernel function according to the form and principle of the combinatorial feature of correlation.In the form of kernel function,the combinatorial feature of correlation is constructed implicitly,which avoids the curse of dimensionality.And the combinatorial feature of correlation is extended to higher order polynomial form by the correlation kernel function.The correlation kernel function is essentially an improved polynomial kernel function.In this paper,the correlation kernel is applied to SVM classification algorithm,and PCC-SVM algorithm is proposed.The comparison between SVM with polynomial kernel function and PCC-SVM is carried out on public datasets.The results show that the classification effect of PCC-SVM is better than SVM with polynomial kernel.