| With the rapid development of information technology,the cardinality of data increases rapidly.In order to mine valuable information in the data,data mining has received extensive attention.With the growth of the cardinality of data,the dimension of features is also increasing,which caused the problem of “the curse of dimensionality”.Feature selection,as an effective pre-processing steps,can solve this problem.Most of the existing feature selection methods can only handle static data while the existing online stream feature selection methods in this scenario can only handle a single type of data.In addition,the problem of instance ambiguity exists widely in real life,and most of the existing online streaming multi-label feature selection algorithms extended from single-label learning,without fully considering the difference between labels,and many feature selection algorithms rely on specific parameter.Rough set is an important model in granular computing theory.In the process of data processing,no relevant domain knowledge is required,and potential rules can be automatically obtained from the data.Therefore,based on rough set theory,this paper studies online stream feature selection for single-label hybrid data and online stream feature selection for multi-label data.The main work of this paper is as follows:1.In the scenario of online stream features,a hybrid neighborhood granulation method is proposed for hybrid data,and a method is constructed from the perspectives of the dependence between features and labels and the correlation between features.A linear fusion feature importance measurement method based on granular computing is constructed.Based on this,the importance and redundancy of features are analyzed,and an online streaming feature selection method for hybrid data is proposed.Experiments are constructed on 9 real datasets,using feature subset size,running time,and prediction accuracy as evaluation metrics.On the premise of ensuring time consumption and feature subset size,this method achieves the best results.The experimental results verify the effectiveness and feasibility of the proposed method.2.In the scenario of online stream feature,for specific parameter problems in existing algorithms,we propose an adaptive neighborhood granulation method.Based on the domain granulation method,a feature importance measurement method for multi-label data is proposed,considering the difference in importance between labels.And then,a multi-label online streaming feature selection algorithm based on adaptive neighborhood is proposed.We constructed a set of experiments on 4 Mulan real datasets with 5 mainstream methods on 7 evaluation indicators in three aspects: feature subset size,time,and classification performance.The experimental results verify the effectiveness and feasibility of the method. |
