Research On Key Techniques Of Partial Multi-label Learning Algorithm

In the framework of multi-label learning,each object is represented by an instance which is associated with multiple labels.The learning goal of multi-label learning is to assign multiple appropriate labels to unknown instances.In traditional multi-label learn-ing research,a common assumption is that each training example is accurately labeled with all relevant labels.However,in many real world applications,this assumption hard-ly holds because the precise annotation is usually difficult and costly.Instead,annotators may roughly assign each instance a set of candidate labels.In addition to the relevan-t labels,the candidate label set also contains some irrelevant labels.We formalize this learning problem as a new framework called Partial Multi-label Learning(PML).Com-pared with the multi-label learning problem which requires accurate annotation,partial multi-label learning is more suitable for practical application.How to use this kind of data to build an effective prediction model is a hot topic in multi-label learning.This paper tries to solve the partial multi-label learning problem,and the main contributions are summarized as follows:1.We propose a novel partial multi-label learning by low-rank and sparse decom-position method.The proposed algorithm firstly utilizes the idea of low-rank and sparse decomposition to capture the ground-truth label matrix and irrelevant label matrix from the observed label matrix.Secondly,our method introduces a trace norm regularization to constrain the ground-truth label matrix by considering the correlation information among the ground-truth labels.Then,the proposed method presents an l1-norm regularization to constrain the noisy label matrix by assuming that the irrelevant labels are sparse.Fi-nally,the prediction model is learned by using the learned ground-truth label matrix and the observed feature information.In order to reduce the complexity of the prediction model,the feature mapping matrix is constrained via trace norm regularization.The pro-posed method enables simultaneously capturing the ground-truth label matrix from the observed label matrix and learning the prediction model via low-rank and sparse decom-position scheme.Extensive experimental results validate the effectiveness of our method.2.We present a novel global-local label correlation method for partial multi-label learning,which mitigates the effect of the noisy labels by making full use of both global-local label joint correlations,and trains the predictors simultaneously.On one hand,to better learn the global label relationship,we construct a label coefficient matrix inspired by the low rank representation scheme,which helps to remove the noise and recover the relevant labels.On the other hand,given the fact that the related tags will encourage their corresponding classifier outputs closer,to better learn the local label relationship,we introduce a new label manifold regularizer to preserve the local label manifold structure,which helps to promote learning the label prediction models.In addition,We also present an l1-norm regularization to constrain the noisy labels to be sparse,since the outlier is usually sparse in real-world applications.By jointly taking advantage of the global and local label correlations,our proposed algorithm achieves superior performance on both the synthetic and real-world data sets from diverse domains.3.We introduce a novel partial multi-label learning with noisy side information approach.For reducing the influence of outliers in feature matrix,the proposed method decomposes the observed feature matrix into an ideal feature matrix and an outlier feature matrix by using the low-rank and sparse decomposition scheme.The ideal feature matrix is constrained to be low rank by considering that the noise-free feature information always lies in a low dimensional subspace.The outlier feature matrix is assumed to be sparse by considering that the outliers are usually sparse among the observed feature matrix.In addition,our approach introduces an ideal label confidence matrix and constrains the confidence matrix to keep the intrinsic structure among feature vectors by utilizing graph Laplacian regularization.Our method succeeds in reducing the influence of noisy features and avoiding the negative effect of redundant labels,which makes our method more robust and applicable than previous partial multi-label learning algorithm.Experimental results further show that the proposed method is more effective than other partial multi-label learning algorithms.4.We propose a novel tensorized subspace representation method for incomplete multi-view partial multi-label learning,in order to reconstruct full representations for in-complete view data,and meanwhile train the predictor for unseen samples.Specifically,we utilize the Low-Rank Representation(LRR)scheme to obtain the subspace represen-tation for recovering the incomplete-view data.Moreover,to further complete the missing information in the view,we integrate together all the subspace representations of different views via tensorized subspace representation,which captures both the global relationship of all views and explores the correlations within each view.In addition,the indicator matrix is introduced to constrain the entries of reconstruct matrix to be equal to original data if they are not missing.Finally,we learn the multi-view partial multi-label learning classifier by utilizing the low rank and sparse decomposition from the recovery multi-view data.To the best of our knowledge,the proposed method is the first MVPML work of jointly considering incomplete views and redundant labels within one framework.Ex-perimental results conducted on fifteen multi-view partial multi-label learning data sets show that the proposed algorithm achieves the best performance in comparison with six state-of-the-art multi-label learning methods.4.We propose a novel tensorized subspace representation method for incomplete multi-view partial multi-label learning.We firstly utilize the Low-Rank Representation(LRR)scheme to obtain the subspace representation for recovering the incomplete-view data.Moreover,to further complete the missing information in the view,we integrate together all the subspace representations of different views via tensorized subspace rep-resentation,which captures both the global relationship of all views and explores the correlations within each view.In addition,the indicator matrix is introduced to constrain the entries of reconstruct matrix to be equal to original data if they are not missing.Fi-nally,we learn the multi-view partial multi-label learning classifier by utilizing the low rank and sparse decomposition from the recovery multi-view data.To the best of our knowledge,the proposed method is the first multi-view partial multi-label learning work of jointly considering incomplete views and redundant labels within one framework.Ex-perimental results conducted on fifteen multi-view partial multi-label learning data sets show that the proposed algorithm achieves the best performance in comparison with six state-of-the-art multi-label learning methods.