Estimating the threshold for maximizing expected gain in supervised discrete Bayesian classification
Robert S. Lynch, Peter Willett · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
When mining discrete data to train supervised discrete Bayesian classifiers, it is often of interest to determine the best threshold setting for maximizing performance. In this work, we utilize a discrete Bayesian classification model, and a gain function, to determine the best threshold setting for a given number of training data under each class. Results are demonstrated for simulated data by plotting the expected gain versus threshold settings for different numbers of discrete training data. In general, it is shown that the expected gain reaches a maximum at a certain threshold. Further, this maximum point varies with the overall quantization of the data. Additional results are also shown for different gain functions on the decision variable.