A new model for classifying DNA code inspired by neural networks and FSA
PKAW'06 Proceedings of the 9th Pacific Rim Knowledge Acquisition international conference on Advances in Knowledge Acquisition and Management
| Hi-index | 0.00 |
One of the major concerns with supervised learningapproaches to text classification is that they require alarge number of labeled examples to achieve a high levelof effectiveness. Labeling such a large number ofexamples poses a considerable burden on human experts.Two common approaches to reduce the amount of labeledexamples required are: (1) selecting informativeuncertain examples for human-labeling and (2) usingmany inexpensive unlabeled data with a small number oflabeled examples. While previous work in textclassification focused only on one approach, weinvestigate a framework to combine both approaches insimilarity-based text classification. By applying our newthresholding strategy (RinSCut) to uncertainty sampling,we propose a new framework which automatically selectsinformative uncertain data that should be presented tohuman expert for labeling and positive-certain data thatare directly used for learning without human-labeling.With our similarity-based learning algorithm (KAN),experiments have been conducted on Reuters-21578 dataset. Our proposed scheme has been compared withrandom sampling and previous conventional uncertaintysampling, based on micro and macro-averaged F1. Theresults showed that if both macro and micro-averagedmeasures are concerned, the optimal choice might be ourframework.