Class Imbalance Multi-label Classification with Common and Label Specific Features
Multi-label classification mainly deals with the problem that instances data is associated with multiple class labels.Most of the existing multi-label methods use the same data representation consisting of all features to distinguish all labels.However,due to the different characteristics of each label,unified features cannot fully differentiate them,which brings negative effects and increases time cost to model training.Therefore,it becomes a challenge to improve the model classification performance by utilizing the most discriminative features for each label.In addition,the problem of class imbalance in reality can also result in a decline in the performance of multi-label learning models.Motivated by this,we propose a new approach of class imbalance multi-label classification with common and label specific features.Firstly,we find the nearest neighbors of seed instances,and then use interpolation techniques to obtain the features of synthetic instances to solve the problem of class imbalance.Secondly,in order to find the most representative features for each label,we introduce l1-norm and l2,1-norm regularizers constraint coefficient matrix to extract label-specific features and common features.Finally,we use label correlation to achieve similar model output of associated labels,and instance correlation to ensure that associated features share corresponding label distribution information to improve classification performance.Extensive experiments show a competitive performance of proposed method against other multi-label learning approaches.
万方数据
维普
