Accelerated Multi-granularity Reduction Based On Neighborhood Rough Set

The core goal of feature selection algorithm is to distinguish irrelevant or redundant features in high-dimensional data sets.Attribute reduction algorithm based on rough set theory is one of the effective methods to deal with such dimension reduction problems.Neighborhood rough set is a model that can effectively process data with continuous values.It can deal with feature extraction in complex data by changing the size of neighborhood radius and constructing multi-granularity structure.The existing researches on neighborhood rough sets mainly focus on the background of single granularity,which makes it an inefficient parameter optimization process to explore the appropriate neighborhood radius and lacks the information sharing and fusion mechanism between different granularity.In order to make up the above deficiencies,this paper studies the multi-granularity attribute reduction based on neighborhood rough set,and introduces two different multigranularity attribute reduction frameworks.By analyzing the correlation between adjacent granule,this method realizes a fast algorithm of attribute reduction and reduces the search space of traditional multi-granularity attribute reduction algorithms.However,this method can not effectively synthesize multi-granularity information,and there are still some redundant calculations.In this paper,an efficient multi-granularity attribute reduction structure is designed.It can reduce the calculation time of multi-granularity reduction to the level of single-granularity reduction.By using voting strategy,attribute evaluation under multiple granularity is realized,and the optimal attribute is always kept in the reduction,which effectively improves the reduction quality.The comprehensive experiment based on 12 UCI data sets shows that,compared with other multi-granularity attribute reduction algorithms,our method not only generates higher quality of multi-granularity attribute reduction,but also has the best time efficiency.