Neighborhood Adaptive Rough Set Model For Multi-label Data

In a multi-label dataset,each sample can be associated with more than one class label.For the classification task,there may be redundant attributes in a large number of attributes of high-dimensional multi-label data.These redundant attributes will increase the complexity and training cost of the classification model,and affect the classification effect to some extent.Attribute reduction based on rough sets is one of the effective methods for dimensional reduction of high-dimensional data.Its main idea is to delete redundant attributes and obtain conditional attribute sets which are more important to the classification task without reducing the classification ability.As an important extension of classical rough set model,neighborhood rough set is an important method for dimensional reduction of high-dimensional data with real values,which is widely used in data mining,pattern recognition and other fields.multi-label neighborhood rough set is an important extension of neighborhood rough set for multi-label data.However,the current model does not reflect the adaptability to complex data distribution;In addition,the effect of label correlation on attribute reduction is not considered.To solve the above problems,this paper aims to develop a more effective multi-label neighborhood adaptive rough set model.The main research contents and conclusions are as follows:(1)Multi-label neighborhood adaptive rough set modelThe current multi-label neighborhood rough set model needs to set a uniform neighborhood radius threshold artificially,which is subjective and does not consider the distribution characteristics of samples.In this paper,the concept of attribute standard deviation of samples relative to labels is proposed,and the neighborhood relation based on this concept is constructed.A multi-label neighborhood adaptive rough set model and the attribute reduction algorithm Adapt-ARML are established.The model can adaptively determine the neighborhood threshold for each sample according to the distribution information of the sample.The validity of the proposed model and algorithm is verified on five numerical multi-label classification datasets.(2)Attribute reduction based on label correlationIn multi-label data,labels are not completely independent of each other.Mining the correlation between labels and using it to improve the performance of attribute reduction algorithms is a problem worth exploring.This paper proposes an attribute reduction framework based on label correlation.Firstly,the framework uses label frequent item sets to mine the correlation between labels and finds the subset of labels with strong correlation within the group.On this basis,two attribute reduction algorithms based on global label correlation and local label correlation are proposed from the perspective of label space and sample space,making full use of label correlation.The effectiveness of the proposed reduction algorithm is verified on five numerical multi-label classification datasets.