InterActive Feature Selection

We study the effects of feature selection and human feedback on features in active learning settings. Our experiments on a variety of text categorization tasks indicate that there is significant potential in improving classifier performance by feature reweighting, beyond that achieved via selective sampling alone (standard active learning) if we have access to an oracle that can point to the important (most predictive) features. Consistent with previous findings, we find that feature selection based on the labeled training set has little effect. But our experiments on human subjects indicate that human feedback on feature relevance can identify a sufficient proportion (65%) of the most relevant features. Furthermore, these experiments show that feature labeling takes much less (about 1/5th) time than document labeling. We propose an algorithm that interleaves labeling features and documents which significantly accelerates active learning.