Research On Error Correcting Output Codes With Maximum Margin

Multi-class classification is an important research field in machine learning,however,it is hard to achieve a good performance by solving it directly.Therefore,it is a simple and general strategy to decompose the multi-class problem into a set of binary problems.Error Correcting Output Codes(ECOC)is a general framework for decomposing the multi-class problem into a set of binary problems.ECOC is usually divided into two steps:encoding and decoding.Considering the traditional encoding strategies usually is data-independent,we propose an encoding strategy based on maximum margin.Furthermore,Considering the sep-arability of the binary problems will influence the performance of ECOC,we further propose a subclass maximum margin tree error correcting output codes.Finally,beyond encoding and decoding,we propose a subclass fine-grained ECOC framework.This framework adds splitting and filtering steps beyond traditional ECOC,as a result,ECOC can solve complex and linearly non-separable problems better.The main contributions of this paper are as follows:(1)Maximum margin tree error correcting output codes(M~2ECOC)is proposed.M~2ECOC uses a data-driven approach to measure the separability among classes.Concretely,M~2ECOC uses maximum margin to evaluate the separability among classes and guides the construction of the coding matrix.By this way,the coding matrix is more representative and ECOC can obtain better classification performance.(2)Subclass maximum margin tree error correcting output codes(SM~2ECOC)is proposed.Based on M~2ECOC,SM~2ECOC further considers that the separability of binary problems can directly affect the performance of ECOC.When the problem is linearly non-separable,SM~2ECOC splits the original complex class into fine-grained subclasses through clustering algorithms.Therefore,it is easier to construct simple and linearly separable problems.(3)Subclass fine-grained error correcting output codes(SFG-ECOC)is proposed.On the basis of SM~2ECOC,SFG-ECOC adds splitting and filtering steps beyond encoding and decod-ing.Before encoding,in order to solve the complex and linearly non-separable problems better,SFG-ECOC splits the original complex class into fine-grained subclasses through clustering al-gorithms.Before decoding,the most representative classifiers are automatically selected,which can not only reduce the complexity of the model,but also effectively improve the classification performance.