Research On Imbalanced Time-Series Multi-Classification

Mining time series data and imbalanced data are two serious problems in the field of data mining,while the problem of imbalanced time series classification includes both the research difficulties,being a comprehensive research hotspot.In this paper,we focus on the problem of imbalanced time-series multi-classification,in which minority classes and majority classes are both more than one class.The contributions in this article are as follows:(1)take imbalanced time-series multi-classification as a whole research topicInstead of solving problems by divided into three subproblems,which are time series classification,imbalanced classification and multi-classification,the paper studies imbalanced time-series multi-classification as a whole research topic.The paper emphasizes the relationship between multi-classification and imbalance problem in imbalanced time-series multi-classification,providing a new idea for solution.(2)propose a joint training framework of imbalanced time-series multi-classificationThe paper first analyzes problem in combination of financial field,formalizes related concept and describes the problem in mathematics.Based on the fact of multi-classification generating imbalance problem,it brings up the idea of solving imbalance problem with multi-classification and put forward "bi-classification equilibrium" method to split multiple classification problems.In this paper,a joint training model for imbalanced time series is established: training classifiers for minority classes and majority classes separately,then combining two classifiers with serial classification or parallel classification and classification threshold,effectively solving imbalance problem in multi-classification.(3)establish a multi-classification joint training model based on serial classificationThis paper analyzes and establishes a multi-classification joint training model under serial classification,and clarifies process framework and evaluation indexes.Choosing two representative time series classification methods as general classifiers for the joint training model,the paper dived into related basic theory,established time series classification algorithm based on LSTM and KNN algorithm based on DTW and analyzed with research topic.(4)apply the multi-classification joint training model to financial fieldAs a case study,the paper applies the multi-classification joint training model based on serial classification to financial field.First of all,analyze the experimental data in detail and select the two representative algorithms in third chapter as the multi-classification joint training model algorithm and contrast algorithm.Then carry out three imbalanced behavior classification experiments,which prove the validity of the joint training framework of imbalanced time-series multi-classification.