Research On Incremental Mechanisms Of Feature Selection And Robust Fuzzy Rough Computing Models For Ordered Data

In the real world,all walks of life are faced with massive amounts of ordered data,that is,there is an order or preference relation in the data.These ordered data have the characteristics of large volume,high dimensionality,noise,and real-time dynamic evolution,which bring great challenges to ordered data mining and analysis,knowledge acquisition,and decisionmaking.Therefore,how to conduct effective and efficient mining and knowledge discovery for ordered data has become one of the research hotspots in the field of information science.Feature selection,as a common data dimensionality reduction method,can remove irrelevant or redundant features and reduce the amount of data,thereby solving the problem of “curse of dimensionality”in the learning model and improving the performance of the learning model.Incremental learning is a key technology for processing data with dynamic characteristics.This technology can continuously learn new knowledge from newly arrived samples on the basis of existing knowledge,so as to achieve the purpose of efficient knowledge acquisition.Rough set theory provides an effective mathematical tool for dealing with inconsistent and uncertain information in data analysis,which is a completely data-driven approach and does not require any prior knowledge of data.For data containing noise,constructing a robust rough set model is a feasible method to deal with the uncertainty problem of data containing noise.This thesis takes the ordered decision system as the research object,adopts dominance-based rough set theory and incremental learning method,mainly studies the efficient and effective feature selection method and the robust fuzzy rough calculation model for the ordered decision system in a dynamic environment.The specific research works are as follows:(1)For dynamic ordered decision systems,the incremental feature selection approaches of object-oriented set change are investigated.First,we take dominance condition entropy as a feature evaluation index,and combine with heuristic reduction strategy to construct a heuristic feature selection algorithm based on dominance condition entropy.Then,the dominance relation matrix and the dominance diagonal matrix are defined,and the matrix calculation method of the dominance conditional entropy is designed.Finally,for the dynamic ordered decision system with the variation of object set,the incremental update mechanisms of dominance conditional entropy are studied,and then efficient feature subset update algorithms are designed.The experimental results prove that the proposed incremental feature selection algorithms improve the computational efficiency.(2)For the ordered decision system containing attribute noise,the robust fuzzy rough calculation model and the incremental feature selection approach are investigated,respectively.First,combined with the idea of neighborhood,the traditional fuzzy dominance relation is improved to ameliorate its fault tolerance for attribute noise.On this basis,a robust fuzzy dominance rough set model,i.e.,the fuzzy dominance neighborhood rough set model is established.Then,the fuzzy dominance neighborhood conditional entropy is defined as a feature evaluation index,its monotonicity is analyzed,and then a non-monotonic heuristic feature selection algorithm based on the index is proposed.Finally,the matrix calculation method of the feature evaluation index is designed,and the update mechanism of the matrix calculation of the index is designed.On this basis,efficient feature subset update algorithms for dynamic ordered decision systems are proposed.The experimental results show that the proposed feature evaluation index has good robustness and the incremental feature selection algorithms are efficient.(3)For the monotonic classification tasks containing label noise,β-precision fuzzy dominance neighborhood rough set and a feature selection method considering feature interaction are studied,respectively.First,the reason why the traditional fuzzy dominance rough set model is sensitive to label noise is analyzed.Then,combining the β-precision quasi-Tnorm and the β-precision quasi-T-conorm,a β-precision fuzzy dominance neighborhood rough set model is established to combat label noise.Under the framework of the proposed model,a series of information entropy based uncertainty measures are extended,the interaction between features is studied,and then the feature evaluation index of dynamic weight change is constructed.Finally,a self-adaptive weighted interaction feature selection algorithm is designed.The experimental results show that the proposed model has good robustness and the proposed feature selection algorithm has good performance.(4)For the monotonic classification tasks containing label noise,the soft fuzzy dominance rough set and the feature selection method considering multiple correlations are studied,respectively.First,aiming at the shortcoming of traditional fuzzy dominance rough set is not robust to label noise,a new robust model is proposed,namely the soft fuzzy dominance rough set model,and the anti-noise mechanism of the proposed model is explained in detail.Then,on the basis of the proposed model,the multiple correlations between features are explored,including redundancy,interaction,and complementarity,and then a feature evaluation index is constructed.Finally,a feature selection algorithm considering multiple correlations is designed.The experimental results prove that the proposed model has better robustness and the proposed feature selection algorithm has better performance.