Transfer Learning Method For Few-label And Zero-label Classification Tasks

Classification technology is a hot topic in data mining,which is widely used in various fields of social life.A classification model is obtained by learning,and each attribute set is mapped to a predefined class label.With the development of deep learning and neural network,classification task also ushered in a new development,but it is still faced with the problems of difficult to obtain annotation data and difficult annotation work.For the situation of few labels,due to the lack of sufficient annotation data,the models trained by conventional classification methods have unstable classification effects and the generalization performances are poor.While in some special fields,even the annotation data can not be obtained,at this time,due to the lack of label information to supervise model for training,the conventional classification methods are no longer applicable.Transfer learning can make use of the similarities of data,model,task and so on,and apply the knowledge learned from a large number of labeled source domain to the target domain to help complete the task of target domain classification.At present,for the situation of few labels or even no label,instance-based transfer learning approaches make data distributions of source domain and target domain similar through instance weighting strategy.The disadvantage of these approaches are that it is difficult to determine the value of weight.Feature-based transfer learning approaches find the common feature representations between source domain and target domain,and use these features to transfer knowledge.But theses approaches are easy to damage internal structures of domains.Considering that parameters of the model can reflect the learned knowledge,parameter-based transfer learning approaches apply parameters of the model trained by source domain data to target domain.However,these model parameters are not easy to converge.Relational-knowledge-based transfer learning approaches transfer analogy relationship,which are less used and only applicable to the relationship domain.In view of above problems,the innovative achievements of this paper are as follows:(1)A parameter transfer method based on nonnegative sparsity constraint and discriminant joint probability is proposed for few-label classification task.A nonnegativity-constraint sparse stacked denoising autoencoder is constructed to improve the data reconstruction ability and robustness of the model which is used to extract more effective high-order features.A method based on discriminative joint probability to measure discrepancy of two domains is designed to reduce the discrepancy of domains while retaining the category information,which is helpful for classification training.In addition,a parameter transfer method based on dynamic selection is designed,determining whether the parameter is set to freeze according to the sensitivity.The simulation results show the superiority and innovation of the proposed method.In the task of few-label classification,this method can make full use of a small amount of labeled data in target domain and has a good classification performance.(2)This paper proposes a transfer learning method based on bi-directional cycle generative adversarial network for the classification task when no label is available.In this paper,a bi-directional cycle generative adversarial network is constructed to transform data distributions in two directions,which makes full use of the information of source domain and target domain.The generated intermediate domains play the role of data enhancement.In addition,a cross-domain alignment method based on class-wise discrepancy of domains is proposed to prevent the loss of class information in the process of cross-domain distributions alignment.The simulation results show the superiority and innovation of the proposed method.In the task of zero-label classification,this method can match data distributions of source domain and target domain as well as possible while preserving their internal structures,and has a good classification performance.