Research On Feature Selection Method Based On Rough Set Theory And Its Applications

Rough set theory,as a tool and method in intelligent information processing,can effectively help us discover useful knowledge and information from dynamic,massive and complex data,so that it has received widespread attention in the field of feature selection.Furthermore,the feature selection method based on rough set theory can not only effectively analyze static data scenes,but also process dynamic data scenes,which makes this method a key part of data dimensionality reduction and plays an important role in data mining.However,with the in-depth research on the feature selection method based on rough set theory,it is found that there are still some problems that need further exploration.For example,the forward heuristic positive region reduction method in static data scenarios is an exemplification of the feature selection method.When selecting candidate attributes,it does not consider the existence of multiple conditional attributes with the highest importance and ignores the influence of decision attributes on the relevance measurement between condition attributes,so that the obtained reduction set has weak generalization ability.In addition,the online streaming feature selection method in the dynamic data scenario,the threshold calculation method under the high relevance attribute evaluation criterion in this method is susceptible to the influence of the attribute dependence numerical data distribution.Meanwhile,the attributes that do not meet the evaluation criterion will be directly remove,and the subsequent work is no longer used,which causes some important attribute information to be lost,and the classification accuracy of the obtained attribute subset is low.To this end,related research on the above problems are conducted in this thesis:In view of the problems existing in the positive region reduction method based on forward heuristic search strategy,information granularity is first used in this thesis,and an optimization strategy for candidate attribute selection is put forward.Secondly,the concept of interaction information in information theory is used to eliminate redundant attributes by calculating the interaction information between attributes.On this basis,an attribute reduction algorithm based on information granularity and interaction information is proposed.Finally,comparative experiments are conducted on the highdimensional microarray gene expression data sets.The experimental results show that the proposed algorithm can obtain reduction results with strong generalization ability and high classification accuracy.In view of the problems existing in current online streaming feature selection methods,from the perspective of attribute dependence,an optimization method for threshold calculation under high relevance attribute evaluation criterion based on the set of conditional attributes selected at the previous moment in this thesis.Secondly,according to the obtained threshold,the newly arrived attributes are classified into highand low-relevance attributes,and different execution strategies are designed for different attributes.On this basis,an online streaming feature selection method based on different attribute execution strategies is proposed.Finally,comparative experiments are conducted on multiple data sets.The experimental results show that the proposed method can obtain a subset of attributes with higher classification accuracy.In summary,in the feature selection method based on rough set theory,different optimization algorithms from static and dynamic data scenarios are proposed in the thesis.And its purpose is to obtain an attribute set with strong generalization ability and high classification accuracy in the data dimensionality reduction stage of data mining,thereby effectively improving the efficiency of data processing.