Thresholding Strategies for Text Classifiers: TREC-2005 Biomedical Triage Task Experiments
Luo Si, Tapas Kanungo · Text REtrieval Conference · 2005
We participated in the triage task of biomedical documents in the TREC genomic track. In this paper we describe the methods we developed for the four triage 1 subtasks. Logistic regression and support vector machine algorithms were first trained to generate ranked lists of test documents. Then a subset of the test documents was identified as positive instances by selecting the top-k documents of the ranked lists. Deciding on the ideal value for k requires a good thresholding strategy. In this paper we first describe two thresholding strategies based on i) logistic regression and ii) support vector machines. In addition to these methods, we describe a thresholding method that combines the outputs from logistic regression and support vector machine by applying a joint thresholding strategy.