Research On Multi-objective Attribute Reduction Based On Decision Rough Set Model

With the arrival of the era of big data,the vast amounts of information generated by the Internet contain a large amount of uncertainty or incomplete information.The decision-theoretic rough set model is a mathematical tool of dealing with uncertain and imprecise data.It extends the rough set theory by combining the Bayes decision theory and rough set theory.And,it can be seen as a rough set model with fault tolerance and high interpretability.As one of the most important concepts in decision-theoretic rough set model or other rough set models,attribute reduction can quickly find and delete redundant attributes from the vast amounts of information,which reduces the data dimensions and improves the efficiency of dealing with information.However,most attribute reduction methods in decision-theoretic rough set are single objective reduction,and the results of different attribute reduction methods are all different.That is to say,the single objective attribute reduction methods may prefer a certain criterion and obtain a biased result,which will make a difficulty for users choosing the most appropriate attribute reduction methods.To handle this problem,this thesis proposes a multi-objective optimization attribute reduct model by combining several attribute reduct criteria.Based on the proposed model,this thesis also proposes two kinds of multi-objective attribute reduction algorithms.The main contents of this thesis are as follows:First,the multi-objective optimization attribute reduct model and the corresponding multi-objective attribute reduction algorithm.In general,multi-objective optimization problem is always used to deal with the problem of multiple conflicting or competing objectives.In this thesis,a multi-objective optimization attribute reduct model is proposed,which considers three kinds of criteria as sub objective functions,including the size of positive region,the decision cost and the mutual information.In the corresponding algorithm,the basic framework of the proposed attribute reduction method is similar to NSGA-II(Non-Dominated Sort in Genetic Algorithm),and it also considers the wrapper method of feature selection.The comparative experiments verified the feasibility of multi-objective optimization attribute reduct model and the efficiency of the corresponding algorithm.Second,the ensemble learning attribute reduction algorithm.The main idea of ensemble learning is to generate a better learning model by considering the integration of multiple learners.Based on the proposed multi-objective attribute reduct model and the idea of ensemble learning,an ensemble learning attribute reduction algorithm is proposed by considering the correlation between attributes and labels.The comparative experiments based on the proposed algorithm also designed,which verified the efficiency of ensemble learning attribute reduction algorithm.Third,application of text classification based on multi-objective attribute reduction.In this part of thesis,the data set is the Chinese text data set.The above proposed multi-objective optimization attribute reduction algorithm and ensemble learning attribute reduction algorithm are used to deal with the text data,and the results of these two algorithms are compared with original text data without any processing.The results of the comparative experiments verified the necessity of attribute reduction on text data and the efficiency of these two proposed algorithm.The proposed algorithm in this paper can obtain the fewer attributes of data,and improves the accuracy of classification with less misclassification cost.