The Study Of Robust Semi-Supervised Classification Algorithm Based On Label Prediction And Propagation

In the community of machine learning,label propagation(LP),as a kind of popular semi-supervised data classification methods,has attracted widespread attention due to its effectiveness and efficiency than those traditional classificartion methods.However,existing LP algorithms still suffer from several drawbacks that may decrease the classification results potentially.First,the classification process is usually sensitive to the noise and redundant information in input data,i.e.,the robust properties of existing LP methods are not strong,which will decrease the label prediction ability.Second,traditional LP methods explicitly separate the graph construction from the label prediction process,while the pre-calcualted weights cannot be ensured to be optimal for subsequent label estimation,which may also suffer from the tricky issues of selecting optimal neighborhood size or kernel width.Third,most existing LP methods focus on processing the single-view samples,without considering multiple-view data,which may lead to the decreased classification performance in reality.Four,most existing methods performs label propagation based on the original input space and original soft label space.But the original input space and soft label space usually have various noise and unfavorable features that may result in inaccurate label prediction results directly.To address the aforementioned shortcomings,we propose some novel innovative strategies in this thesis.Extensive results over real-world image databases have verified the effectiveness of our methods.The major contributions are summarized as follows:(1)We propose a new transductive label propagation method,termed Robust Adaptive Label Propagation by Double Matrix Decomposition.To improve the classification result by addressing the issues caused by the noise,outliers and redundant information in the original input space and soft label space,we propose an effective strategy to remove the mixed signs and noise from the predicted soft labels by the matrix decomposition.That is,our method further decomposes the predicted soft label matrix into a clean soft label matrix and a noise term,and then estimates the hard label based on the recovered clean soft labels for more accurate classification.In addition,our method integrates the graph construction with the LP stage cleatly,which can ensure the obtained weights to be promising for classification.(2)In order to further improve the classification accuracy based on the double matrix decomposition,we propose a novel framework termed Robust Triple Matrix Recovery based Auto-weighted LP for Classification,which considers recoverying the weighting space as well.That is,the proposed method recovers the underlying clean data space,clean label space and clean weighting space jointly by clearly decomposing the original data,predicted soft labels or weights into a clean part plus an error part fitting noise.In addition,our method also integrates the auto-weighting by minimizing the joint reconstruction errors based on the recovered clean data and clean label to explicitly ensure the weights to be joint optimal for both data representation and classification.By classifying data in the recovered clean label and clean weight space,the label prediction results can be potentially improved.(3)In real applications,a sample may have multiple-view representations in various subspaces.To extend LP from the single-view to multiple-view scenario,we also propose an Adaptive Multiple-view Label Propagation framework for semi-supervised classification.The proposed method aims at enhancing the performance by discovering the effects of multiple-views rather than on the single-view.Besides,our method integrates the multiview LP and adaptive multiple graph weight learning into a unified framework,which can exploit the complementation using a linear transformation to make different adaptive weights form different view spaces.By mining valuable knowledge from different views in an adaptive manner,the tricky issue of selecting neighborhood size or kernel width can be avoided.