This paper addresses the problem of actively constructing a training set for the linear model with sparse structure. This problem usually occurs in the scenario that no nonlinear mappings give similar performance for large-scale learning data, but it has to train a linear model quickly. In this paper, an active framework is proposed to reduce the time expense further in constructing the training set. The training examples are iteratively selected by matching partial components and their weights given by the classifier in pairs, in order to exploit model's sparsity to precisely separate out more informative examples from others in a short time. The proposed framework is evaluated on a group of classification tasks, including the texts and images.
展开▼