Batch Mode Sparse Active Learning

Sparse representation, due to its clear and powerful insight deep into the structure of data, has seen a recent surge of interest in the classification community. Based on this, a family of reliable classification methods have been proposed. On the other hand, obtaining sufficiently labeled training data has long been a challenging problem, thus considerable research has been done regarding active selection of instances to be labeled. In our work, we will present a novel unified framework, i.e. BMSAL(Batch Mode Sparse Active Learning). Based on the existing sparse family of classifiers, we define rigorously the corresponding BMSAL family and explore their shared properties, most importantly (approximate) sub modularity. We focus on the feasibility and reliability of the BMSAL family: The first one inspires us to optimize the algorithms and conduct experiments comparing with state-of-the-art methods, for reliability, we give error-bounded algorithms, as well as detailed logical deductions and empirical tests for applying sparse in non-linear data sets.