Research On Data-driven Multi-criteria Decision Making Methods Based On Multiple Classifier System

To better meet the demands of decision making for accuracy and objectivity,various data-driven multi-criteria decision making(MCDM)studies have been widely performed in recent years.Among these studies,the prevalent way is to design a decision parameter self-learning mechanism to improve or extend the traditional MCDM approaches.Although this way can ensure the applicability of MCDM in the big data environment,it is still difficult to satisfy the actual demands for decision accuracy due to the over-simple internal structures of traditional MCDM approaches.To address this issue,data-driven MCDM approaches based on multiple classifier system are investigated in this thesis based on the analysis of different MCDM processes,which aims to enhance the decision accuracy by using the learning capability of multiple classifier system on the condition that the decision results are understandable.The main contents of this thesis are listed as follows:(1)Focusing on preference learning in the context of MCDM,this thesis proposes a data-driven preference learning approach based on dynamic ensemble of multiple classifiers.In the proposed approach,the interval numbers are firstly used to model the preference learning problem in the context of multiple criteria,and then a dynamic ensemble mechanism is designed based on ensemble error and diversity of multiple classifiers.The designed ensemble mechanism allows us to dynamically adjust the classifier weight in the weighted ensemble process according to the competency difference of base classifiers when predicting the overall assessments of different alternatives.Afterward,a linear MCDM optimization model is constructed to learn a set of criterion weights to transform the weighted ensemble predictions of multiple classifiers,which ensures the comprehension of the final preference recommendations for alternatives.Large amounts of ultrasound grading diagnosis data of thyroid nodules are collected to examine the performance of the proposed approach,and experimental results indicate that the proposed approach can effectively satisfy the demand for high decision accuracy on the condition that the decision recommendations are understandable.In addition,the experiments based on thirty public classification datasets show that the designed dynamic ensemble mechanism also outperforms existing mainstream ensemble learning approaches.(2)Focusing on the individual decision making in the context of MCDM,this thesis proposes a data-driven decision-making approach based on dynamic selection of multiple classifiers.In the proposed approach,the interval numbers are firstly used to model the individual decision-making problem,and then a dynamic selection mechanism is designed based on data similarity and classification accuracy of base classifiers.The designed selection mechanism enables us to select the best base classifier for each alternative to be evaluated.Afterward,a weight learning model is constructed to learn a set of representative criterion weights to transform the prediction of the gold standard derived from the selected base classifier,which ensures the comprehension of the final gold-standard recommendations of alternatives.Lots of diagnostic data of thyroid nodules with pathological findings are collected to examine the performance of the proposed approach,and experimental results indicate that the proposed approach can effectively fit the mapping relationship between the individual assessments of historical alternatives on criteria and corresponding gold-standard assessments,and ensure the whole decision performance.In addition,the experiments based on ten public binary classification datasets show that the designed dynamic selection mechanism can also perform as well as existing mechanisms without subjectively determining the size of the neighbor region.(3)Focusing on the group decision making in the context of MCDM,this thesis proposes a data-driven group decision making approach based on dynamic ensemble selection of multiple classifiers.The proposed approach first uses interval numbers to model group decision making problem and then designs a dynamic ensemble selection mechanism based on the predictive performance of the optimal individual base classifier.The designed ensemble selection mechanism enables us to select the best subset of base classifiers for each alternative to be evaluated.Afterward,a decision optimization model is constructed based on cross-entropy to learn a set of group criterion weights to improve the acceptance of group experts for the group decision recommendation results.The experimental study based on ultrasound diagnosis data of multiple radiologists and the corresponding pathological findings shows that the proposed approach can not only effectively learn the common decision preferences of group experts but also effectively help handle the group decision making with largescale alternatives.In addition,the experiments based on thirty public classification datasets indicate that when the error-independence assumption is not met in practice,the dynamic ensemble selection mechanism designed in this thesis can achieve better classification performance than existing dynamic ensemble selection approaches due to the consideration of the ensemble performance of different classifiers in the ensemble selection process.The research results of this thesis mainly have two contributions.On the one hand,it can further promote the development and improvement of decision-making theory in the big data environment,and on the other hand,it can provide effective references for constructing Artificial Intelligence-based data-driven management decision-making framework.