A New Classifier Based On Density Ratio Model And Kernel Method

In imbalanced classification tasks,the low detection rate of the minority often makes us suffer a huge loss.In order to improve the performance of classifier in im-balanced learning,this paper proposed a new classifier based on density ratio model and kernel method:density ratio classifier(DRC).The kernelization of the density ra-tio model,the efficient algorithm and the discrimination rule based on the estimation of ROC curves are indispensable for DRC dealing with imbalanced classification tasks effectively.In the second chapter,this paper first introduces the density ratio model and its related theoretical basis.In order to estimate the real density ratio functions in complex cases well,the kernel method is used to extend the density ratio model to the kernel density ratio model(KDRM).Secondly,the penalized empirical likelihood method is combined with the lower bound algorithm and Nystrom method to solve KDRM ef-ficiently,which greatly reduces the calculation burden.Lastly,the discriminant rule based on the estimation of ROC curves enables DRC can effectively solve the difficult imbalanced classification tasks by threshold moving.In Chapter 3,this paper uses the characteristic analysis to obtain the convergence rate of the log density ratio estimation under certain conditions.In addition,from the perspective of hypothesis testing,this paper explains that the ROC curve corresponding to density ratio is the best curve in ROC space.And the excellence of DRC is discussed as well.The fourth chapter contains simulation research and analysis of real datasets.The simulation study tests two methods to change the imbalance ratio,and gives the exper-imental results of binary and multi classification cases.The actual data sets also take into account the data complexity and imbalance.The analysis of performance of the classifiers from two aspects shows that DRC outperforms other classifiers in imbalance problems.Finally,in the summary of the paper,it points out problems and corresponding solutions,as well as the next direction of research.