Missing Label Feature Selection via Global Similarity of Samples and Relief
The original Relief model can only analyze complete single-label data,and the relevant improved models have not considered the global similarity between samples.Therefore,a missing label feature selection method is designed based on the global similarity between samples and the Relief model.Firstly,in order to com-plete the missing labels corresponding to the samples,all samples are divided into missing sets and complete sets under each label.By calculating the Euclidean distance between samples,the nearest neighbors of the missing la-bel samples in the complete set are searched,thereby proposing a label completion strategy to supplement the missing labels of the samples.Secondly,to measure the similarity relationship of different samples in the global space,the cosine similarity function is used to calculate the feature similarity between samples,and the label simi-larity between samples is calculated based on the overlap degree of the sample label sets,combining the above two kinds of similarity to construct the global similarity between samples.Then,in order to determine the same-class neighbors and different-class neighbors of the target sample in the multi-label decision system,the same-dif-ferent class discrimination relationship between samples is defined based on the global similarity between the tar-get sample and the rest of the samples.Finally,a new feature weight iteration formula is constructed based on the improved Relief model,and a missing label feature selection algorithm based on sample global similarity and Re-lief is designed.The classification performance of the proposed algorithm is analyzed and tested on 8 multi-label datasets,and the experimental results show that the proposed algorithm is effective.
feature selectionmulti-label learningmissing labelRelief model
NETL
NSTL
万方数据
维普
