Research On Several Issues In Weakly Supervised Classification

For classification with strong supervision,each instance is associated with a unique and valid label for learning a mapping function from feature space to label space.However,many real-world classification tasks cannot meet the strict requirements of strong supervision information,such as:(1)an instance is associated with multiple labels,(2)an instance is associated with an erroneous label.To adapt to these practical conditions,two weakly supervised classification frameworks have been proposed,i.e.multi-label classification and noisy label classification.In multi-label classification,each instance can be associated with multiple labels for learning a mapping function from feature space to label power set.To improve the learning performance of multi-label classification system,an effective way is to manipulate the feature space to help construct multi-label classification model.For instance,the label-specific features strategy derives specific feature representation for each label,on which a classifier is built.Many existing methods based on label-specific features decouple the stage of label-specific features generation from that of classification model induction.However,this two-stage decoupling strategy may lead to suboptimal generalization performance.In this paper,a wrapped method(WRAP)is proposed to wrapping the two stages by(kernelized)linear model with empirical loss minimization and pairwise label correlation regularization,which generates label-specific features for each label in the embedded feature space and learns a(kernelized)linear model simultaneously.Comparative experiments on sixteen benchmark datasets show that the wrapped label-specific features method can effectively improve the generalization performance of multilabel classification system.In noisy label classification,the dataset consists of a portion of noisy samples with erroneous labels.To deal with them,the small loss trick is used by most methods to screen out possible clean samples.Furthermore,to prevent ignoring useful samples with large loss,some hybrid methods adopt sophisticated schemes to integrate other technologies to achieve better classification performance.In this paper,an ensemble method(RLME)is proposed to yield a simpler but more robust hybrid method,which constructs several pairs of models by randomly bipartitioning the dataset evenly and calculates sample weights by the distance between the prediction distribution and the given label.In addition,RLME introduces Mixup and label masking for data augmentation and semi-supervised learning,so as to improve the consistency and the robustness of the ensemble model.The comparative experiments on several real-world datasets fully show the effectiveness of RLME to deal with label noise with significant better classification performance than other methods.