Research On Classification Learning Method For Missing Data Markers

With the continuous progress of science and technology,a large amount of data is produced in all fields of the world at any time and anywhere.These data are unmarked and lack of marking information,while manual marking is time-consuming and laborious.Therefore,in order to solve the problem of insufficient data marking,semi-supervised learning and transfer learning came into being,and gradually became an important content in the field of machine learning research.Semi-supervised learning uses both labeled data and unlabeled data to train models to predict target data.Migration learning first uses the labeled source domain data to learn the model,and then uses the learned model to predict the unlabeled target domain data.Although there are many methods to classify data,there are still the following shortcomings: 1)Manifold regularization(MR)is a classical framework in semi-supervised learning,but the smoothness constraint of MR is realized for all pairs of samples,that is,each pair of samples is regarded as an object,and the smoothness of a single sample is not considered.2)In migration learning,Domain Adversary Neural Networks(DANN)is a typical algorithm to solve unsupervised classification learning problem,which is easy to produce negative migration in the process of aligning the whole data distribution of source domain and target domain.In addition,only one classifier trained by labeled samples in the source domain is used to predict the target data,and the result may not be reliable.Therefore,in order to solve the above problems,this paper mainly includes the following two contents:Firstly,in order to reduce the errors that may occur in pairs of samples while still retaining the point-by-point smoothness of samples,and considering the smoothness of each sample instead of pairs of samples,a new method for semi-supervised image classification is proposed,namely Point Wise Manifold Regularization(PW?MR).This method not only considers the smoothness of each sample,but also introduces the local density to measure the importance of each sample,thus promoting the prediction accuracy of MR framework.Secondly,in order to make full use of the knowledge of labeled samples,a bidirectional discriminant domain adaptive network(BDDAN)is proposed.In this method,domain anti-neural network(DANN)is used for feature alignment to reduce domain spacing,and domain invariant features are further learned by image transformation.In this process,two classifiers are learned at the same time,and the consistency loss function is introduced through the probability outputs of the two classifiers,thus improving the reliability of prediction.In addition,the labeled samples are used to calculate the similarity between classes,so that the target samples predicted to be of the same class have similar soft labels,which helps to improve the classification performance.At last,experiments are carried out on data sets,and the experimental results show that the proposed PW?MR and BDDAN methods can improve the classification performance and improve the accuracy of prediction.