Information Granulation-based Decision-theoretic Rough Set Models And Methods

With rapid development of internet technology,modern sensor technology and other emerging technologies,diverse data types has become the main form of data in data-driven intelligent decision making.Multipleattribute decision making widely exists in the social life field as an important research content of decision science.Due to the ambiguity and uncertainty of decision objects,decision-theoretic rough set theory gives a view for solving multi-attribute decision making problems as a new method with characteristics of noise tolerance and risk cost sensitivity to deal with uncertain decision problems.However,in real complex decision making problems,decision data often contain multi-view,multi-level/multi-scale and high dimensions information,and the decision task also presents complex characteristics of multi-level and multigranulation.As a decision method with single granular structure,classical decision-theoretic rough set can not cope with complex decision making problem solving.Granular computing is a discipline used to study thinking mode and problem solving based on multi-level granular structure.Multigranulation,multi-view and multi-level description,reasoning and problem solving for various practical problems are the main research contents of granular computing.By introducing human's multigranulation cognitive mechanism to data analysis of decision-theoretic rough sets,decision-theoretic rough set models and decision making methods are constructed from new viewpoints.The main contributions obtained are as follows:(1)Through referencing dynamic granulation cognitive ability,we have constructed a new probabilistic rough set framework under dynamic granulation,called decision-theoretic rough sets under dynamic granulation;The related rough decision making methods on the view of dynamic granulation are developed,which provides new ideas and methods for complex problem solving efficiently.In order to solve problems of inefficiency that exist in attribute reduction and rule acquisition algorithms caused by the non-monotonicity of probabilistic positive regions in decision-theoretic rough sets,we introduce a new probabilistic rough set approximation approach through combining the local rough set principle and the idea of dynamic granulation,in which the monotonicity of decision positive regions is satisfied.An efficient rough feature selection method is developed,which realizes an efficient feature selection process through monotonic change of decision positive regions under dynamic granulation.Theory analysis and experimental results show that the computational efficiency of the algorithm is certainly improved and alleviated the overfitting phenomenon;We construct three-way decision models under dynamic granulation combined with the three-way decision theory,and a two-stage dynamic three-way decision-rule acquisition algorithm is designed.Experiments illustrate effectiveness and efficiency of the algorithm.(2)We have developed multigranulation decision-theoretic rough set models and methods based on probability fusion,which provides a feasible method for the analysis of a kind of multisource data analysis with risk.Information fusion of multi-source data is a common problem in the field of data analysis,focusing on which a concrete multigranulation information fusion method is presented from viewpoints of probability theory based on the Bayesian decision under multiple granular spaces,that converts the information fusion to the probability fusion with easier to express.Based on different fusion strategies,we have constructed optimistic/pessimistic multigranulation decision-theoretic rough set models respectively;By introducing a concept of the approximate distribution reduction to the proposed model,the granular structure selection problem under multiple granular spaces is investigated.Aiming at the inconsistency of decision making in multigranulation probability rough approximation,a new concept of approximate distribution quality is defined to describe the lower approximate distribution of target decision.A granular structure reduction algorithm with keeping ?-lower approximate distribution unchanged is proposed.(3)By introducing multi-level solving methods to decisiontheoretic rough set models,we have developed decision-theoretic rough set models and methods for multi-scale data,which provides the reference method for people to find the decision making knowledge with the cost-sensitive characteristics from the multi-scale data.The traditional multi-scale decomposition makes decision data for each scale too coarse or thin,we propose the concept of generalization scale,which makes multi-scale data decomposition more flexible in the sense of the generalization.Based on Bayesian decision theory,we have constructed multi-scale decision-theoretic rough set model in the new definition of multi-scale generalization decision table,in which characteristics of cost-sensitive and noise tolerance are fully considered.That is more suitable for the real multi-scale data decision problem solving;Considering the non-monotonicity of probabilistic decision regions on multi-scale granular layers,there are some difficulties in the theory and semantic interpretation for usual optimal scale selection methods.By introducing cost-sensitive learning,a new optimal scale selection algorithm is designed,in which the optimal scale selection problem is transformed into a cost optimization problem.Experimental results verify the effectiveness and interpretability of the proposed algorithm.In this thesis,by introducing granular computing methods for multigranulation solving into decision-theoretic rough set models from the viewpoints of multi-view,multi-scale and dynamic granulation,we have established a rough decision making model and methods system based on the information granularity.In the view of granular computing,a new way of the rough decision problem solving is explored,that has a certain theoretical significance and application prospect for the real complex decision problems.