Multi-view Multi-label Learning Based On Dimension Reduction Of Label Space

With the development of technology,the label space of the sample is increasing.This puts computational stress on the classic multi-label learning model and reduces the performance of multi-label learning.Researchers are inspired by the dimension reduction of feature space,and put forward the idea of dimension reduction of label space.With the advent of the artificial intelligence era,samples or tasks are no longer simply described with single view.When the sample has multiple perspectives and the label space dimension of the sample is large,the existing method of multi-label learning and multi-view multi-label learning based on label space dimension reduction cannot be solved.On this basis,this paper proposes two methods for multi-view multi-label learning based on label space dimension reduction.The main content and motivation points of this article are listed below:This paper first proposes an integrated multi-view label space reduction method.The method calculates a corresponding predictive label space on the feature space of each view of the multi-view,and uses the error between the prediction of each view and the real label space to learn a series of weight to measure its proximity to the real label.In this paper,experiments on three typical multi-label image datasets prove that the proposed method is more accurate than the results of voting results and the optimal feature space.On the basis of the first work,a complete multi-view multi-label learning algorithm based on label space reduction is proposed.The algorithm uses the Hilbert-Schmidt independence criterion to balance the consistency and complementarity between multiple views based on maximizing the relationship among label space,latent space and feature space.The implicit spaces are related to each other by iterative updating.After obtaining the accurate latent space,the feature space of each view is embedded into the corresponding latent space and the predicted result is obtained by using the decoding matrix and voting.Experimental results on six different sized multi-labeled image data demonstrate that the proposed method is superior to the existing method based on label space reduction strategy.