Research On Weakly Supervised Class Imbalance Multi-label Classification

Traditional multi-label learning methods assume that all relevant labels of each training example are available.However,in practice,multi-label learning involves a large number of labels and complex semantics,making it difficult and time-consuming to obtain sufficient high-quality supervision for many tasks.Consequently,multi-label datasets used for training often contain weakly supervised information with missing and noisy labels due to various factors.Moreover,the class imbalance problem inherent in multi-label data is considered to be a significant factor that harms the performance of machine learning algorithms,particularly when dealing with missing or noisy data.In view of the above research difficulties,this thesis proposes several weakly supervised imbalanced multi-label classification algorithms.Weak multi-label learning with missing labels via instance granular discrimination.Considering that in practical applications,there are often missing relevant labels,which leads weak performance of multi-label learning models.In this thesis,we propose two co-training multi-label methods,C2 ML and C2 MLE,where C2 ML is based on the direct linear mapping and C2 MLE is based on ensemble learning.In this method,the feature structure and label manifold are used to recover the incomplete label matrix and train the classification model simultaneously.In order to alleviate the risk of imbalance caused by the sparse labeling problem,an adaptive penalty factor is imposed on the deviation prediction of different labels.In addition,the instance granularity discrimination is introduced into the framework to characterize the approximate distribution structure of the data.Finally,matrix vectorization,Concave-Convex Programming(CCCP)and block coordinate descent techniques are used to solve the proposed optimization problem.Extensive experiments show the superiority of the proposed method compared with some state-of-the-art methods.Multi-label classification with weak labels by learning label correlation and label regularization.Considering that the real-world objects usually exhibit more sophisticated properties,such as abundant irrelevant features,incomplete labels,noisy labels,as well as class imbalance.In this thesis,we propose an integrated multi-label learning framework ML-INC that trains the multi-label model while addressing the aforementioned issues simultaneously.Firstly,we decompose the observed label matrix into an incomplete ground-true label matrix and a noisy label matrix by employing the low-rank and sparse decomposition scheme.Secondly,high-order label correlation and label consistency are used to learn a label confidence matrix to recover missing labels.In addition,the low-rank assumption is used to help to obtain the label correlation.Thirdly,a label regularization term is introduced to alleviate the effects of class imbalance in the label matrix,and a sparsity constraint is imposed on the feature mapping matrix to select relevant discriminative features.Finally,Alternating Direction Multiplier method(ADMM)was used to deal with the optimization problem,and comprehensive experiments were carried out to verify the effectiveness of the method.