Abstract

Although semi-supervised, multi-view algorithms have been successfully used to reduce the need for labeled data, our experiments indicate that existing approaches are sensitive to the levels of correlation and incompatibility of the views in a domain. To cope with this problem, we introduce a new multi-view algorithm, Co-Test (Co-EM), which interleaves active learning (Co-Testing) and semisupervised learning (Co-EM). Co-Test (Co-EM) has a robust behavior on a large spectrum of problems because of its ability to ask for the labels of the most ambiguous examples, which compensates for the weaknesses of the underlying semisupervised algorithm. We compare Co-Test (Co-EM) with four semi-supervised algorithms on a parameterized family of 60 text categorization problems in which we control the level of view correlation and incompatibility. The experiments show that Co-Test (Co-EM) clearly outperforms the other algorithms without using more labeled data; these results are further confirmed by two additional experiments performed on real-world datasets.

Date

January 1, 1970

Authors

Ion Muslea, Steve Minton, Craig A Knoblock

Journal

IJCAI-01 Workshop on Text Learning: Beyond Supervision