SELF-WEIGHTED DEEP SUBSPACE CLUSTERING WITH FUZZY LABELS

. In weakly-supervised learning, the background knowledge of applications could be given in the form of fuzzy labels, where experts are not sure about the exact class annotations for all data objects (e.g., images), but they can provide candidate labels for a subset of the objects. Subspace clustering aims to segment data objects into different clusters, each of which is related to a subspace. In this study, we propose a novel self-weighted deep subspace clustering algorithm, named DSCF, to learn the subspaces of datasets provided with a small number of fuzzy labels. In the proposed loss function, we consider the local relationship of data objects to reconstruct the input data and design the constraints on the self-weighted similarity of objects to embed the fuzzy labels. The self-expression matrix for final spectral clustering is learned by minimizing the loss function to jointly optimize the reconstruction error and the comprehensive constraints. The experiment results show that the proposed DSCF outperforms the compared algorithms on five gray-scale image datasets. Further experiments on the cardinality of fuzzy labels are conducted and the effectiveness of fuzzy label simulation is discussed.