Research On Unsupervised Domain Adaptation Methods Combined With Pseudolabel Improvement

Unsupervised domain adaptation is an important technique in the field of machine learning that aims to use source domain data to help perform tasks in the target domain without using labeled data from the target domain.Under unsupervised conditions,the target samples are unlabeled and cannot be accurately aligned for conditional distribution.The combination of source domain samples and target samples after feature preprocessing can be implemented for classifier training by generating pseudo-labels.Inaccurate pseudo-labeling will bring unreliable distribution alignment problems and lead to classification errors.In this thesis,we propose a domain adaptation method regarding projection learning and pseudo-label improvement considering sample similarity as well as variability,and the main research is as follows:(1)To address the problem of inaccurate target pseudo-labeling due to insufficient consideration of data structure information of target samples,this thesis proposes a method on feature projection learning combined with sample weighting to improve the pseudo-labeling by combining sample structure information.First,the method minimizes the target classification error by learning feature projection and projecting sample features from the high-dimensional original space onto a common low-dimensional subspace to reduce the cross-domain joint distribution bias.Secondly,the method considers sample similarity using clustering properties.A more accurate pseudo-label is assigned to the target sample by finding the nearest source domain class centroid through nearest neighbor search for the clustering class centroid of the target sample.Finally,the projected source and target domain samples and pseudo-labels are used for sample weighting to train the target classifier to classify and identify the target samples.The method is experimented on the dataset to verify that it can improve the target classification accuracy.(2)To address the problem of classification error accumulation during iterative learning due to the incomplete accuracy of the initial weak classifier in labeling the target samples.In this thesis,a method based on manifold subspace learning and pseudo-label selection is proposed.The method firstly uses manifold learning to preprocess the samples and obtain manifold features by transforming the original samples to the Grassmann manifold space.Next,pseudo-label selection assignments are made for the target samples using nearest class prototypes,structured prediction and class centroid matching to select the pseudo-label with the highest probability for a given target sample.Experiments are conducted on the dataset to verify that the method can improve the classifier performance.In this thesis,a weighted domain adaptation method based on pseudo-label improvement with projection learning and a selective-method based on manifold subspace learning are proposed in the context of unsupervised domain adaptation.Using feature projection to preprocess the source and target domain samples first reduces the domain distribution differences;using the structural information of the target domain data distribution and the similarity of the samples,the problem of pseudo-label labeling error is mitigated and the classification accuracy of the target samples is improved.