Based On Incomplete Supervision Multi-label Classification Algorithm

With the wide application of multi-label learning methods in many fields,the accurate classification of multi-label data has become one of the important topics for studying multi-label learning methods.For multi-label data,the rapid development of the Internet makes it very convenient to obtain unlabeled data,but it takes a lot of time and money to get the labeled multi-label data.In this paper,a multi label classification algorithm based on incomplete supervision is proposed by making full use of a small number of labeled examples and a large number of unlabeled examples.According to the two main ways to realize incomplete supervised learning,this paper divides the proposed algorithm into two parts:multi-label classification algorithm based on active learning and multi-label classification algorithm based on active semi-supervised learning.Based on the active learning multi label classification algorithm,firstly,according to the relationship between multi-labels,an asymmetric multi-label correlation matrix is constructed by using the earth mover's distance;Then,the method of combining the entropy of the binary source and the correlation matrix between multi-label is used to calculate the information content of the example-label pair,and this is used as the active learning sampling standard.Finally,the selected example is given to a human expert for labeling and iteration to complete the active learning process.This algorithm not only considers the inter-relationship between multi-labels,but also considers the information contained in labeled and unlabeled data,which further improves the classification performance of multi-label classifiers.The multi-label classification algorithm based on active semi-supervised learning is based on the basis of active learning multi-label classification algorithms,in order to further improve the classification efficiency of the algorithm,semi-supervised learning is added to automatically select and label unclassified multi-label data.The semi-supervised learning strategies used in the algorithm include classification algorithms based on ordered weighted average operators in fuzzy rough sets,semi-supervised SVM classification algorithms,and semi-supervised k-nearest neighbor classification algorithms.In a single iteration,the algorithm carries out manual annotation of active learning and automatic annotation of semi supervised learning simultaneously,which further improves the classification efficiency of multi-label classifier.The two multi-label classification algorithms based on incomplete supervision proposed in this paper are compared with the other seven multi-label classification algorithms on three multi-label data sets for four evaluation indicators.The experimental results show that the two proposed classification algorithms have further improved classification performance compared with other traditional multi-label classification algorithms and the newly proposed classification algorithms.