Multi-class Cost-sensitive Learning Based On Decision-making Rough Set Model

In recent years,with the development and popularization of computer and Internet technology,the huge amount of data that produced subsequently,is vastly different in scale and complexity from past information,in addition,there are some uncertainty and fuzziness.Decision-theoretic rough set model(DTRS)as an effective method to deal with the imprecise and uncertain and problems,it educes a systematic method to calculate the decision thresholds according to the loss function matrix by introducing the Bayesian decision process,then we can obtain the three-way decisions framework based on the three regions of rough set,and it is a good solution to the problems of how to make a reasonable decision when the information is not enough for some users.Many existing studies have been finished on the basis of the classical two-class DTRS model.For multi-class classification problems,most of them are transformed into multiple binary classification problems and can be solved by the classification method of two-class DTRS,the classification process requires users to provide more loss functions and reduces the efficiency of calculation.In view of this,this paper combines DTRS with cost-sensitive learning and proposes a multi-class decision-theoretic rough set model.The cost-sensitive learning is studied based on the multi-class decision-theoretic rough set model,studies of this paper are covered from the following aspects:Firstly,the extension of DTRS in the classification model.Loss function matrix plays an important role in DTRS,the decision thresholds can be easily calculated by the matrix in binary classification problems.In this paper,we consider loss functions as the object of studying,the DTRS is combined with cost-sensitive learning,and the loss functions of multi-class classification problems can be derived from a cost matrix of classical cost-sensitive learning.A multi-class decision-theoretic rough set model is proposed and a cost-sensitive three-way decisions classification algorithm based on the model is presented.The effectiveness of the proposed model and algorithm in dealing with multi-class classification problems is analyzed by comparing the experimental results.Secondly,a multi-phase cost-sensitive learning method based on the multi-class decision-theoretic rough set model.The output of DTRS classification method is a three-way decisions result,only objects in the positive region can be determined the class label with a high confidence.However,objects in the boundary region are made a delayed decision due to lack of information and objects in the negative region are made a rejected decision because of the low confidence,that is,objects in the boundary and negative regions cannot be determined the class label.In order to solve above problem,in this paper,a multi-phase cost-sensitive learning method based on the multi-class decision,theoretic rough set model is proposed,the boundary and negative:regions can be eliminated by using classification process of multiple phases and the three-way decisions results are finally transformed into the two-way decisions classification results.The experimental results indicate that the proposed algorithm can get a better.classification performance than standard classifier and some cost-sensitive learning methods.Thirdly,text classification based on the multi-class decision-theoretic rough set model.Text classification is a popular research topic in recent years.In this paper,the Sogou Chinese Text is chosen as a corpus to train a text classifier based on the multi-phase cost-sensitive learning algorithm,and compare it with several commonly used machine learning classification algorithms,the comparison.results show that our algorithm can obtain a higher classification precision,a high classification recall and a lower classification cost in text classification.The results further highlight the cost sensitivity of the proposed algorithm and extend the application of decision-theoretic rough set in some practical problems.