Unsupervised Active Learning via Subspace Learning

Unsupervised active learning has been an research topic in machine community, with the purpose of choosing representative samples to be labelled unsupervised manner. Previous works usually take minimization data reconstruction loss as criterion select which can better approximate original inputs. However, are often drawn from low-dimensional subspaces embedded arbitrary high-dimensional space many scenarios, thus it might severely bring noise if attempting precisely reconstruct all entries one observation, leading a suboptimal solution. In view this, this paper proposes novel Active Learning model via Subspace Learning, called ALSL. contrast previous approaches, ALSL aims discovery low-rank structures data, and then perform sample selection based on learnt representations. To end, we devise two different strategies propose corresponding formulations under representations respectively. Since proposed involve several non-smooth regularization terms, develop simple but effective optimization procedure solve them. Extensive experiments performed five publicly available datasets, experimental results demonstrate first formulation achieves comparable performance state-of-the-arts, while second significantly outperforms them, achieving 13\% improvement over best baseline at most.