Multi-Granularity Ensemble Classification Algorithm Based on Attribute Representation

In the face of the complex and changeable information systems, in the field of machine learning traditional multi-classification models cannot achieve a dynamic classification process, and it cannot solve some problems such as disease diagnosis. Because some diagnosis procedures are too expensive, it is necessary to judge whether the patient is likely to be ill through some preliminary diagnosis, thereby reducing the cost of the process. Sequential three-way decisions as a multi-granularity classification algorithm, which is used to solve dynamic classification problems in multi-granularity space. The sequential three-way decision model sorts attributes by balancing the cost of decision results and decision process, then a multi-level granularity space is constructed. With the injection of information in turn, objects that meet the conditions are classified at different granularity levels. It can be said that the sequential three-way decision model solves the problem of excessive costs for decision process. Therefore, many scholars at home and abroad have optimized the sequential three-way decision model from perspective of cost-sensitive. However, in some cases, the sequential three-way decision model in the coarse granularity space is prone to decision conflicts, that is, the same object gets multiple different classification results. Therefore, many attributes must be considered in the fine granularity space, which leads to low classification efficiency. Because of the lack of more information and corresponding strategies, the sequential three-way decision model is unable to process the final unclassified objects. Therefore, this paper combines the ideas of ensemble learning and granular computing to propose a multi-granularity ensemble classification algorithm based on attribute representation. Firstly, constructing a classifier by selecting the attribute representatives of each granularity layer to form an ensemble classifier based on attribute representatives. By synthesizing the different opinions of the classifier which is constructed by attribute representation, the generation of decision conflicts in each granularity layer can be effectively reduced. Secondly, the classification opinions of classifiers in the coarse granularity space are retained through the scoring table to reduce the number of attributes that need to be considered in the fine granularity space. The retained score can make the ensemble classifier which is constructed by attribute representation in the fine granularity space to avoid more likely errors, thereby obtaining a more confident classification result. Finally, there may still be cases where some objects are not classified after all the information has been injected, so the "relatively optimal" strategy is adopted, and the decision class with the least objection rate is used as the final classification result of unclassified objects. In order to verify the validity of the model in this paper, the 14 UCI data sets and 6 real data sets which are related to medical diagnosis are used to conduct horizontal and vertical comparison experiments respectively. Among them, the horizontal comparison experiment includes ten popular multi-classification algorithms. Through experiments, the proposed method in this paper has better robustness, classification efficiency and classification performance than the sequential three-way decisions and other machine learning multi-classification algorithms. Moreover, the multi-granularity ensemble classification algorithm based on attribute representation has improved significantly in the real data sets of medical diagnosis. © 2022, Science Press. All right reserved.