Feature Selection Approach Under Laplacian Kernel Weighting

In the challenges of computational intelligence,high-dimensional data sets often have thousands of features,most of which are redundant and irrelevant.In order to solve the problem of high feature dimension and improve the efficiency of pattern classifier,feature(attribute)selection is usually carried out before classification.Rough set theory provides a soft computing method for the approximation of clear sets in the feature selection stage,but it can only deal with symbolic information systems.Because there are a lot of real value data in the information system,the incremental concept of discernibility using fuzzy rough theory is more natural and effective than rough set.In addition,realistic datasets often contain mixed attributes,which means that the attributes in the dataset contain different types of values.In the process of feature selection,uncertain information usually leads to information loss under the rough set theory.First of all,because the traditional fuzzy rough approximation method is easily affected by the noise of a single object,this paper selects the order weighted average operator of Laplace kernel to construct the weight,and gives the fuzzy rough and resolution matrix model based on the Laplace weighted kernel in the context of real value information system.Considering the complementarity of the two measures,the aggregation framework based on the two robustness models is selected to achieve the goal of feature selection.Secondly,based on hybrid information system combining grey relation theory and fuzzy rough sets,based on the calculation model of direct edge boundaries is given and its combined with entropy difference as a description of measurement uncertainty evaluation function to choose,the model is used for rapid processing mixed not complete information system,more suitable for the decision attribute of quantitative information system.Finally,this paper compared the two feature selection methods of real value type and mixed type with other similar methods under three classifiers and four indicators,and verified that both of the above two algorithms can reduce feature dimensions and improve classification accuracy.