Research On Multi-label Classification Based On Pseudo-label Attention Ensembl

The purpose of multi-label classification is to predict multiple labels for each unknown instance,and it is an important topic that has been widely studied in the field of machine learning.Multi-label classification has many typical applications in the real world,including recommender systems,text classification,music/image classification,etc.People have gone from studying single-label classification to multi-label classification that explores the correlation between labels.Although many effective algorithms have been proposed,the following key challenges still exist in multi-label classification research: How to effectively explore the complex correlation between labels;how to deal with the exponential label output space.All these difficulties make multi-label classification research intractable and challenging.The multi-label classification algorithm based on pseudo-label attention ensemble proposed in this paper focuses on two aspects,one is how to reduce the dimension of the explosion-level label output space under the premise of retaining more complete feature information,and the other is how to more efficiently explore correlations between labels.The encoding-decoding mechanism is introduced to encode the original label space into a low-dimensional pseudo-label representation space,where the pseudo-label representation space is the common latent semantics learned from the real labels,and then,by learning the potential correlation between the pseudo-label representation matrix and the ground-truth labels,the predicted full label space is decoded from the pseudo-label matrix.In addition,inspired by ensemble learning,in feature embedding,a multi-layer perceptron followed by an attention network is used as a learner,and multiple learners are combined to map features to a low-dimensional pseudo-label representation space,and then,in order to ensure the diversity and complementarity of multiple learners,two optional diversity regularization terms are introduced to generate two models.A large number of experimental comparisons on multiple benchmark datasets have confirmed that the multi-label classification algorithm proposed in the paper is more efficient and robust.The main contributions of the model proposed in this paper are:1.In order to solve the exponential label output space and reduce the computational cost and memory requirements.The pseudo-label representation is proposed through the encoder-decoder mechanism by a “soft” way,which is used to reduce the label space and capture the potential correlation between tags.2.In order to explore the correlation between multiple labels more efficiently,multiple attention networks are integrated for feature embedding,so that it can make full use of feature information from multiple different views.In addition,the introduction of exponent and hinge diversity regularization to ensure the diversity and complementarity of multiple learners.