Research And Application On Classification Method Based On Neighborhood Relation Fuzzy Rough Set

Feature selection can yield feature subsets which provide information on behalf of the whole data set in the course of data mining, pattern recognition, machine learning,etc.Compared with processing whole data, feature selection can save more time for aforesaid courses and improve their efficiency. Rough set based feature selection method has become a focus of current research in recent years since it is able to cope with imprecise, uncertain and vague information in original data sets.In this paper, we chiefly study theories such as neighborhood rough set, fuzzy rough set, feature selection and parallel data mining architecture, and so on. First of all,Neighborhood relation fuzzy rough set and its feature selection algorithm are respectively proposed. Also, application in medical image classification with the feature selection method is studied. Finally, parallelized feature selection based on neighborhood relation fuzzy rough set is achieved through using Compute Unified Device Architecture. Main works of this paper are as follows:(1) Theories and approaches of neighborhood relation basis fuzzy rough set model. In order to explore the expansion of rough set generalization model in fuzzy environment and secure a more compact classification model, this paper proposes neighborhood relation fuzzy rough set(NR-FRS) and its feature selection method, which can improve classification accuracy and reduce the cost of data processing. The lower and upper approximation belonging to NR-FRS are constructed by employing fuzzified neighborhood relation. Meanwhile, some reasoning and demonstration are provided in fuzzified neighborhood approximation space. Besides, this paper also analyses the dependency in feature subspace and construes the definition of positive region as well as property dependency in fuzzified neighborhood approximation space. Finally,experiments are carried out on UCI standard data sets. Compare with neighborhood rough set feature selection method, property quantity obtained by NR-FRS feature selection algorithm varies more stably according to parameter values, furthermore,promotion on average classification accuracy can reach 5.2% in the best case.(2) Application of NR-FRS feature selection method in mammography image classification. In order to put such feature selection into practice, this paper has appliedNR-FRS feature selection method to classification on mammography feature data sets that derives from mammography image standard data sets(Mammography Image Analysis Society, MIAS). Firstly, we process images with image processing approaches and extract texture features from them to construct feature data set. Finally, property subset is generated via NR-FRS based selection method and regarded as input for RBF-Support Vector Machine. Such results shows that NR-FRS reduction algorithm has obtained an 82.16% classification accuracy in the best case, which is respectively 21.1%and 27.2% higher than Forward Attribute Reduct method based on Neighborhood Rough Set and Kernel Principal Component Analysis method.(3) CUDA realization of parallelized feature selection algorithm based on NR-FRS executed on mammography image data set. Since there exist some computing intensive tasks among feature selection and classification on mammography data sets, this paper conduct parallel data mining on large scale medical image data set. We employ the well-known parallelism idea called Compute Unified Device Architecture to deal with feature selection on medical images. Specifically, two tasks, data standardization and neighborhood granule compute, are parallelized. Moreover, computing efficiency on different processor(CPU and GPUs) is compared and influences on computing time caused by threads organization are also studied. Experimental results show that computing time has been improved entirely and wholly over such tasks.