Active learning for text classification: Using the LSI Subspace Signature Model

Supervised learning methods rely on large sets of labeled training examples. However, large training sets are rare and making them is expensive. In this research, Latent Semantic Indexing Subspace Signature Model (LSISSM) is applied to labeling for active learning of unstructured text. Based on Singular Value Decomposition (SVD), LSISSM represents terms and documents as semantic signatures by the distribution of their local statistical contribution across the top-ranking LSI latent dimensions after dimension reduction. When utilized to an unlabeled text corpus, LSISSM finds the most important samples and terms according to their global statistical contribution ranking in the corresponding LSI subspaces without prior knowledge of labels or dependency to model-loss functions of the classifiers. These sample subsets also effectively maintain the sampling distribution of the whole corpus. Furthermore, tests demonstrate that the sample subsets with the optimized term subsets substantially improve the learning accuracy across three standard classifiers.