Robust Multi-Label Learning With Missing Label

In multi-label learning,each object represents a single example and is associated with multiple labels at the same time.His initial research mainly focused on the ambiguity in text classification,and has been applied to many fields such as image recognition,speech recognition,and gene classification.Existing multi-label learning algorithms usually assume that the information labeled in the label set is complete.However,in the era of big data,due to the complicated process of data set labeling,the ignorance of some labeling information due to different concerns of labelers,and the lack of professional knowledge required by labelers themselves,it is difficult to collect a complete label data set,and these missing label information will have a huge impact on the construction of the model,not only causing the model to lose its ability to discriminate in these missing categories,but also affecting the prediction performance of the known categories.This problem of complete lack of information on certain labeled categories is different from the problem of local label missing on some known categories in common multi-label learning,and it is impossible to label information through some commonly used methods such as low-rank structure and matrix completion.restore.In order to mine the label information that is completely missing in some categories,this paper proposes an algorithm called MLLHL(Multi-Label Learning with Hidden Labels)that can not only recover the label information that is completely missing in some categories,but also For unknown examples,both missing markers and observed markers are predicted.This article first constructs a complete mark set consisting of marked categories and missing categories.The marked set of marks can be obtained through the complete mark set through the corresponding column selection step.Then,in order to maintain the structural relationship between similar examples in the feature space in this complete marker set,corresponding manifold constraints are added to the constructed complete marker matrix.Secondly,this paper learns a classification model between the feature space and the learned complete label space,and adds corresponding constraints to the model coefficients to maintain the correlation between the labels and avoid the interference of noise data.Finally,for any unknown example,the learned classification model can predict its labeled information on its labeled and missing categories.The experimental results prove that the MLLHL algorithm has achieved better performance than other similar excellent algorithms.And these completely missing category markers can further improve the prediction performance of the model on known markers.In order to verify the practical application effect of the proposed algorithm,this paper applies the MLLHL algorithm to some original text data sets and achieves good experimental performance.