Term-Centric Active Learning for Naive Bayes Document Classification
Sunghwan Sohn, Donald C. Comeau, Won Ho Kim, W. John Wilbur · The Open Information Systems Journal · 2009
In real world document classification, a subset of documents often needs to be chosen for labeling as a training set for a machine learner. Random sampling is generally not the most effective approach for choosing documents to be la- beled. Active learning selects useful examples for labeling to improve the efficiency of learning. We consider two factors in order to measure the usefulness of a document for labeling. Such a document should be 1) largely unknown to the cur- rent learner 2) influential by being close to many other documents. These factors are stated from a document-centric viewpoint. A similar analysis can be made from a term-centric viewpoint. It is the purpose of this paper to present this term-centric approach to active learning using a naive Bayes classifier. We study both document-centric and our new term-centric active learning methods. We find good performance of the term-centric methods on numerous data sets with different characteristics. In addition, a genetic algorithm is employed to compare our results with estimated optimal per- formance at fixed training set size and our results are between 84% and 99% of the estimated optimum.