The Research Of Multi-label K-nearest Neighbor Classification Based On Coupled Similarity

With the development of our society,multi-label classification has gradually become the hot spot of people,and more and more fields(such as scene classification,automatic text categorization,gene functional analysis of bio-informatics,multimedia analysis and so on)have involved in this problem.At present,there are a lot of methods solving multi-label classification,for example BR?BP-MLL?ML-kNN etc.These approaches assume that the data attributes and labels are independent and identically distributed,and neglect its correlation.However,in the real life,there are more or less explicit or implicit connections between data.Considering the correlation between data in multi-label classification can mine the potential information of data,and improve the classification performance.Multi-label k-nearest neighbor is a lazy learning approach,and it can predict the labels of unseen instances by learning the training sets which have known labels.However,this approach does not consider the correlation of attributes and labels.So,in this paper,the explicit(or implicit)correlation of attributes and labels have been taken into multi-label k-nearest neighbor approach,which called a multi-label classification based on coupled similarity(i.e.CSML-kNN).In order to avoid the unscalability of CSML-kNN,this paper reduces the dimension by using feature extraction and feature selection,and proposes an approach based on principal component analysis for multi-label classification(PML-kNN),an approach based on coupled similarity of principal component analysis for multi-label classification(PCSML-kNN),and an approach based on feature selection for multi-label classification(RCSML-kNN).At last,based on the real multi-label data(emotions and yeast),experiments are conducted by ML-kNN?CSML-kNN?PML-kNN?PCSML-kNN and RCSML-kNN.Experiments show that multi-label classification considering feature selection and coupling can not only mine the latent information of multi-label data,but also can predict the label of unknown instances effectively.