Comparative Study On Classifier Evaluation Measures MCC,CEN And ACC

Due to machine learning and pattern recognition is widely application in real life.the classification problem as its core part has been favored by the majority of research scholars pay much attention to evaluate the classification accuracy of classifiers and compare different,is the last stage of the classifier design,and is one of the most critical step in the process of data mining the good performance of a classifier is to determine the classifier could be used in the classified situation of the most important premise,therefore,evaluate whether the classifier performance good degree can be accepted is the key step in the process of classification.Among the classifier evaluation indicators,the accuracy rate ACC(Accuracy)is a classic classifier evaluation index.Due to its logical evaluation method and simple and easy to expand characteristics,it is widely used in various types and problems by various industries.Classification evaluation.Matthews Correlation Coeffic ient(MCC)evaluation index is widely used in the field of bioinfonnatics.It was first used to solve the classification evaluation problem of the second type of unbalanced data sets,and has been extended to the evaluation of many types of problems.Convergence Entropy(CEN)is a classifier evaluation index proposed in recent years that is directly defined on multi-class problems.It fully considers the misclassification information of samples,shows the separation of categories and samples,and has strong identification for classification results.These three excellent classifier evaluation indicators each have their own characteristics.This paper conducts an in-depth comparative study on MCC,CEN and ACC evaluation indicators.First,eight aliasing matrices with the same ACC and four MCCs were constructed and their CENs were calculated.It was found that these CEN values were different.That is to say,when ACC and MCC can't evaluate different classification results,CEN can distinguish them.Secondly,three kinds of problems are taken as examples to make common evaluations of CEN and MCC,CEN and ACC,ACC and MCC.The comparison of the consistency and discriminant degree of the advantages and disadvantages of the two methods shows that the CEN has a high degree of consistency and a certain degree of discriminability compared with the MCC and ACC evaluation indexes.Therefore,the CEN classifier evaluation index has a better evaluation effect on the classifier performance;finally,12 benchmark data sets(including 8 multi-class data sets and 4 second-class data sets)and 6 common classifiers in UCI(The classification results obtained by the two classifiers and four multi-class classifiers are used to analyze the performance of the classifier evaluation index CEN,MCC and ACC,which further proves that CEN has stronger discrimination.In this paper,by comparing the performance evaluation of CEN,MCC and ACC,it is concluded that CEN evaluation index is more applicable in the performance evaluation of classifiers.