Research On Regularized Least Squares-Based Safe Semi-Supervised Learning

Semi-Supervised Learning?SSL?,as one side of machine learning both with labeled data and unlabeled data,aims to achieve better performance by mining the knowledge of unlabeled instances.In some cases,due to the introduction of unlabeled data,it's happens that SSL's performance may be decrease than supervised learning?SL?without unlabeled data.How to mining the knowledge of unlabeled instances safely are gradually becomes one hotspot of research community.On the basis of summarizing the existing Safe SSL?S3L?methods,this thesis combines the regularized least squares method?RLS?and the Laplacian regularized least squares method?Lap RLS?to introduce the dual learning mechanism and the adaptive estimation strategy of risk degree,proposed two works of S3 L.The study of this thesis as follows:?1?This thesis propose a Du AL Learning-based s Afe Semi-supervised learning?DALLAS?,which sets RLS as primeval model and Collaborative Representation-based Classification?CRC?as dual model,and use those models predicts labels and reconstructs unlabeled instances.After calculates reconstruction error,DALLAS will derive risk degree and perceives one risk degree-based S3 L method.For each unlabeled instance,if the reconstruction error is large,then the risk is high and the prediction approaches to SL.Otherwise,the risk is small and it can be used for S3 L.Finally,we carry experiments in some benchmark datasets of medical dataset and digital recognition dataset and experiments shows DALLAS can reduce the risk of unlabeled instances effectively,especially when the time with only fewer labeled instances,then cause reduction of misclassification risk.?2?Another work of the thesis is an l₁norm-based safe semi-supervised learning method.The current S3 L only considers the risk of unlabeled instances,and doesn't analyze the risk of labeled instances.At the same time,the risk degree of unlabeled instances relies on the artificial design evaluation function and lacks the adaptive mechanism of estimating risk degree.Based on the loss function of l₁norm,the second work perceived an optimization problem of S3 L which can reduce the negative effect of both labeled instances and unlabeled instances simultaneously.Then the problem can be solved by an adaptively estimation of risk degree,by introducing an iterative optimization strategy.Finally,we carry out the experiments on benchmark datasets of image dataset or some other dataset,and achieves 87.8% accuracy in Digit1 dataset,which can verify the effectiveness of our work as expected.