Partial multi-label learning with noisy side information

Abstract

Partial multi-label learning (PML) aims to learn from the training data where each training example is annotated with a candidate label set, among which only a subset is relevant. Despite the success of existing PML approaches, a major drawback of them lies in lacking of robustness to noisy side information. To tackle this problem, we introduce a novel partial multi-label learning with noisy side information approach, which simultaneously removes noisy outliers from the training instances and trains robust partial multi-label classifier for unlabeled instances prediction. Specifically, we first represent the observed sample set as a feature matrix and then decompose it into an ideal feature matrix and an outlier feature matrix by using the low-rank and sparse decomposition scheme, where the former is constrained to be low rank by considering that the noise-free feature information always lies in a low-dimensional subspace and the latter is assumed to be sparse by considering that the outliers are usually sparse among the observed feature matrix. Secondly, we refine an ideal label confidence matrix from the observed label matrix and use the graph Laplacian regularization to constrain the confidence matrix to keep the intrinsic structure among feature vectors. Thirdly, we constrain the feature mapping matrix to be low rank by utilizing the label correlations. Finally, we obtain both the ideal features and ground-truth labels via minimizing the loss function, where the augmented Lagrange multiplier algorithm and quadratic programming are incorporated to solve the optimization problem. Extensive experiments conducted on ten different data sets demonstrate the effectiveness of our proposed method.