Research On Rough Feature Selection Methods For Multi-label Classification

As is known to all,in addition to unstructured,multi-modal and multi-related characters,high-dimendional and multi-label are also the main forms of data complexity.For example,a picture may contain "sunset","cloud" and "tree".A document may contain labels such as "Africa","civil war","oil" and so on.Usually,these multi-label data are still high-dimensional,such as the sequences of genes,pixel points of images,and the feature words of text,etc.The space and time complexity of data analysis algorithms will be greatly increased by the data of high-dimensional,and the multi-label character of data will increase model's complexity.Developing high-dimensional multi-label data feature selection method is important for theory of data modeling,machine learning,and it also has many significant potential application values.Rough set theory is a kind of effective methods to deal with uncertain information,it does not need priori assumptions on data,and use information granule to construct the concept of approximation operators,and implement attribute reduction under the design principles,then construct classifiers by rules or other forms,this method has been widely used in many fields of data analysis and mining.In this paper,the fuzzy rough set model and neighborhood rough set model are developed for multi-label data classification.The main research contents and conclusions are as follows:(1)A multi-label fuzzy rough feature selection method is based on label correlation.The tags of multi-label data are not independent,and the existing methods of rough feature selection for multi-label classification do not consider the correlation between labels,which affects the performance of the built model.In order to solve this problem,the concept of attribute-label matrix is proposed in this paper,which describes the correlation between labels well.Then,the fuzzy rough set model for multi-label data is constructed and the corresponding feature selection algorithm is designed.The validity of the proposed method is verified on the open multi-label dataset.(2)A multi-label neighborhood rough feature selection method is based on double space granulation.Compared with the single label data,the label spatial granular structure of multi-label data can provide more abundant information for data analysis,which is not utilized by the existed research work.Therefore,in this paper,from a more comprehensive perspective,a mixture of the label space and feature space method are proposed for the multi-label data feature selection.Based on the idea of neighborhood rough set,a model of neighborhood rough set based on labeled space particle is constructed,and the corresponding feature selection algorithm for multi-label data's classification is designed as well.The validity of the proposed method is verified on the open multi-tag dataset.