Research On Feature Selection And Classification Method Under Multiple Kernel Fuzzy Rough Set

At the present stage,due to the difference of data sources,data information structure and data statistics,the different information between data attributes cannot be ignored.The existing traditional fuzzy-rough set usually use only one measure to calculate the fuzzy similarity relationships with all attributes.This method ignores the difference information of data attributes,which will directly affect the performance of feature selection algorithms and classification algorithms.In order to solve this problem,the multiple kernel learning be combined with the fuzzy rough set,and multiple kernel-based fuzzy rough set is proposed.Based on the kernel-based fuzzy rough set,the different kernel functions are selected to calculate the fuzzy similarity relation with each attribute.Then,the T-norm operator is used to combine the relations.Finally,the combined kernel is used to replace the original single metric to calculate the fuzzy upper approximation,the fuzzy approximation and the attribute dependency degree.This method considers the difference information of data attributes in fuzzy similarity relations,and makes the calculation of fuzzy similarity relations more scientific.For feature selection task,based on multiple kernel-based fuzzy rough set,multi-kernel based fuzzy rough set particle swarm optimization feature selection algorithm(MPSOFS)is proposed.The algorithm uses fitness function to simultaneously select features and the kernels which are used to calculate dependency degree.The calculation of the fitness value is related to the dependency degree and the number of selected features.The experimental results verify the effectiveness of multiple kernel-based fuzzy rough sets in feature selection area.Meanwhile,the experimental results also prove that MPSOFS algorithm has stronger feature reduction ability and better feature selection.For the classification task,based on multiple kernel-based fuzzy rough set,multiple kernel-based multi-functional nearest-neighbor classification(MMNN)is proposed.The algorithm calculates the fuzzy similarity relationship by using the multiple kernel-based fuzzy rough set,and selects the neighbor samples according to the fuzzy similarity relationship.Then,the fuzzy similarity relations are aggregated with the class membership degrees,and the sample class is judged according to the aggregated results.The effectiveness of multiple kernel-based fuzzy rough sets in the classification field is verified by numerical experiments.At the same time,the experimental results show that MMNN algorithm has better accuracy than the other traditional classification algorithms,and the classification performance is more stable.