Research On Sequential Multi-class Three-way Decision Model

As an extension of the traditional two-way decision,three-way decisions plays an important role in dealing with uncertain information.The third decision-making behavior delay decision is introduced into three-way decisions,which can greatly reduce the loss of direct decision-making when information is not sufficient.Three-way decisions gives a semantic interpretation to three decision-making behaviors.The objects in the positive region represent the decision-making of acceptance,the objects in the negative region represent the decision-making of rejection,and the objects in the boundary region represent the decision-making of delay.The traditional three-way decisions mainly focuses on two-class classification.However,multi-class classification is more common in practical application.For example,in the process of medical diagnosis,doctors should not only diagnose whether patients have diseases,but also diagnose the types of patients' diseases.Therefore,the multi-class classification problem based on three-way decisions has attracted extensive attention of scholars.However,there are two problems when dealing with multi-class classification problems based on three-way decisions: first,for large amounts of high-dimensional data,the efficiency of processing multi-class classification problems based on three-way decisions needs to be improved.Second,there may be decision conflict when processing multi-class classification problems based on three-way decision,that is,objects are divided into the positive region of multiple decision classes.Aiming at the above two issues in multi-class classification decision,the specific work of this thesis is as follows:(1)Aiming at the limitation of a large number of redundant elements in the construction of traditional discernibility matrix in rough sets,this thesis improves the method of constructing discernibility matrix.Before calculating the discernible information of any two objects,determine whether the values of the core attributes are consistent firstly.If not,the corresponding element item is set as ? directly without judging other condition.So that the elements in the matrix are reduced.Based on the discernibility matrix,considering the ratio of each condition attribute to the non-empty element in discernibility matrix and the contribution of each attribute to the distinguished object,the concept of attribute weighted importance is given.An improved discernibility information tree based on attribute weighted importance is proposed to further reduce the redundant elements in the discernibility matrix.Finally,an attribute reduction algorithm based on the improved discernibility information tree is proposed and experiments verify the effectiveness of the algorithm.(2)To handle decision conflict based on three-way decisoins to deal with multi-class classification problems,there are currently two main conflict resolution strategies: one is to resolve decision conflict after decision-making,and the other is to resolve decision conflict before decision-making.Considering that decision conflict is also a kind of uncertainty,and insufficient information is the main cause of decision conflict,this thesis combines sequential approach and gradually adds more detailed information to solve the decision conflict problem during decision-making.This thesis combines this decision conflict resolution method with three multi-class classification methods and three sequential multi-class three-way decisions models are proposed,which are defined as "one-versus-rest" model,"one-versus-one" model and "one-versus-multiple" model.Among them,the multi-class classification strategy adopted by the "one-versus-rest" model is to convert m-class decision into m two-class decisions,and then make three-way decisions for each m two-class decisions.The multi-class classification strategy adopted by the "one-versus-one" model is to select any two classes to form a two-class classification,so it can form m(m-1)two-class classification.Then three-way decisions are made for m(m-1)two-class classification.The multi-class classification strategy adopted by the "one-versus-multiple" model is to directly make three-way decisions for m-class simultaneously.The objects with decision conflict in three models will be processed at the next level.Only the objects that can make clear decisions and have no decision conflict can be classified into corresponding decision classes.Finally,this thesis compares the performance of the three conflict resolution strategies and the proposed three models,and verifies the effectiveness of the three models through experiments.