Initial training data selection for active learning

The crucial issue in many classification applications is how to achieve the best possible classifier with a limited number of labeled training data. Active learning is one method which addresses this issue by selecting the most informative data for training. In this work, we argue that the performance of active learning could be improved through carefully selecting the initial training samples. To confirm our argument, we propose three initial training data selection mechanisms based on fuzzy clustering method: center-based selection, border-based selection and hybrid selection. Center-based selection selects the samples with high degree of membership in each cluster as initial training data. Border-based selection selects the samples around the border between clusters. Hybrid selection is the combination of center-based selection and border-based selection. The effects of them are empirically studied on a set of UCI data sets. Experimental result indicates that, compared with randomly selecting initial training samples, hybrid selection can effectively enhance the performance of active learning.