Active transductive KNN for sparsely labeled text classification

Sparsely labeled classification may exist in many real-world applications and it is more challenging than the problems most existing semi-supervised learning/active learning algorithms considered. In this paper, an active transductive framework is proposed for sparsely labeled text classification. It integrates the advantages of semi-supervised learning and active learning, and employs several techniques to cope with the training data bias and sparsity. A batch mode active learning strategy is used to enhance the performance of semi-supervised learning. The fusion of active learning with rechecking strategy, as well as the employment of common feature extraction technique, makes our framework robust to the training data bias and sparsity. Experimental results on several real data sets show that the proposed classification framework is more effective and efficient for sparsely labeled text classification compared with several state-of-the-art methods.

I. Introduction

Sparsely labeled classification may exist in many real-world applications such as content-based image retrieval, online web-page recommendation, object identification and text categorization, where the abundant unlabeled instances are available but the labeled ones are fairly expensive to obtain since manually labeling the training data for a machine learning algorithm is a tedious and time-consuming process, and even unpractical(e.g., online web-page recommendation). Correspondingly, one important challenge for large-scale text categorization is how to reduce the number of labeled documents that are required for building reliable text classifier.