The Study Of Error Correct Output Codes Algorithm Based On Soft Codes And Variable Length Codes

Error correct output codes(ECOC)is a general framework to solve multiclassclassification problems.It can solve multiclass-classification problems by integrating multiple base binary classifiers.Moreover,by taking advantage of the characteristics of self-correcting codes,it can automatically correct part of errors in the final prediction.ECOC algorithms mainly include two phases:the encoding phase and the decoding phase.The encoding phase generates a coding matrix,which is the core of the ECOC algorithm.The quality of the coding matrix directly determines the performance of the algorithm.The algorithm for generating a coding matrix can be divided into the datadependent encoding and data-independent encoding algorithms.The data-dependent encoding algorithm generates a coding matrix based on data distribution.On the contrary,the data-independent coding matrix only relates to the number of classes in datasets.In general,the data-dependent coding matrix gives better results than the dataindependent coding matrix.The decoding phase will determine the final prediction label based on the prediction results of each base classifier.To generate a coding matrix more suitable for data distributions,two new encoding methods are proposed in this dissertation to adapt data distribution from different perspectives.The main work of this dissertation is as follows:(1)In this dissertation,the theoretical background and relevant research of existing ECOC algorithms are reviewed in detail and some famous ECOC algorithms are introduced in detail.(2)This dissertation presents a novel soft coding ECOC algorithm.The coding matrix is composed of continuous values between[-1,+1]instead of the traditional discrete values of {-1,0,+1}.Soft coding is used to express the tendency of a different class to different groups to improve the adaptability of the algorithm to different data.A fine-tuning strategy is proposed to adjust the coding matrix.By fine-tuning the element values in the coding matrix,the coding matrix can be more adaptive and the prediction accuracy of the base classifier can be improved,and the error of decoding can be reduced.(3)This dissertation proposes a two-stage coding method.Considering that different categories in the same dataset have different separability,this dissertation uses a two-stage coding method to encode easy class and hard class respectively and longer codes will be used for the hard class.(4)The two proposed algorithms are verified on UCI datasets and Microarray datasets.The Friedman test and Nemenyi test are used to verify the difference of proposed algorithms.In the end,we demonstrate the effectiveness of the proposed algorithms from multiple aspects.