The Study Of Robust Label Propagation Algorithms For Semi-Supervised Data Classification

Semi-Supervised learning is a technique that can use both labeled and unlabeled data for learning,so it overcomes the shortcoming of fully supervised learning in reality.Label propagation(LP),as a kind of semi-supervised learning algorithm,has attracted considerable attention in recent years due to its effectiveness and fast computational speed.It is also worth noting that virtually all the existing formulations may suffer from three potential drawbacks.First,these methods are sensitive to noise.Second,they perform label prediction after an independent weight construction process,which cannot ensure the constructed graph weights to be optimal for subsequent label propagation.In addition,the neighborhood information of each data is usually determined by using K-neighborhood or ?-neighborhood,but choosing a proper K or ? is also not easy in reality due to complex contents and distributions of various real data.Third,existing models define the weights and propagate label information based on the original high-dimensional data that usually contains the unfavorable features,irrelevant features,noise and even gross corruptions,but the included noise,corruptions and the unfavorable features may result in inaccurate similarity measure and prediction results directly.To address the aforementioned shortcomings,we propose four novel frameworks.Extensive results over real-world image databases will verified the effectiveness of our methods.The major contributions are summarized as follows:(1)We propose two methods to enhance the robustness to noise and reliability of the distance metric used in the existing LP models: 1)Nuclear-norm based Transductive Label Propagation.To model the neighborhood reconstruction error more reliably,we use the nuclear norm that has been proved to be more robust to noise and more suitable to model the reconstruction error than both L1-norm or Frobenius norm for characterizing the manifold smoothing degree;2)Transductive Classification Robust Linear Neighborhood Propagation.We use the L2,1-norm that is proved to be very robust to noise for characterizing the manifold smoothing term.Since L2,1-norm can also enforce the neighborhood reconstruction error to be sparse in rows,i.e.,entries of some rows are zeros.In addition,to enhance robustness in the process of modeling the difference between the initial labels and predicted ones,we also regularize the weighted L2,1-norm on the label fitting term.(2)We propose a new transductive label propagation method,termed Adaptive Neighborhood Propagation by joint L2,1-norm regularized sparse coding.To make the predicted soft labels more accurate for predicting the labels of samples and to avoid the tricky process of choosing the optimal neighborhood size or kernel width for graph construction,our method seamlessly integrates the sparse coding and neighborhood propagation into a unified framework.That is,the sparse reconstruction error and classification error are combined for simultaneous minimization,which clearly differs from traditional label propagation methods that explicitly separate graph construction and label propagation into independent steps,which may cause inaccurate predictions.(3)We propose a novel adaptive label propagation approach by joint discriminative clustering on manifolds.Our framework seamlessly combines the unsupervised manifold learning,discriminative clustering and adaptive classification into a unified model.For transductive classification,we first perform the joint discriminative K-means clustering and manifold learning to capture the low-dimensional nonlinear manifolds.Then,we construct the adaptive weights over the learnt manifold features.Note that high-dimensional data analysis and processing may benefit from dimensionality reduction or feature learning or manifold learning that can discover the underlying subspace structures of high-dimensional data by learning a dimension-reduced representation.It also noted that manifold learning could preserve the important local or global geometry structures among samples and remove unfavorable features as well as noise effectively.(4)In addition,we also use label propagation algorithms for image segmentation.Experimental results show that the proposed algorithm effectively improves the segmentation and interaction of image segmentation.