Estimating the Class Probability Threshold without Training Data
Ricardo Blanco-Vega, Cèsar Ferri, José Hernández‐Orallo, Marïa José Ramírez-Quintana, Camino de Vera · 2006
In this paper we analyse three different techniques to establish an optimal-cost class threshold when training data is not available. One technique is directly derived from the definition of cost, a second one is derived from a ranking of estimated probabilities and the third one is based on ROC analysis. We analyse the approaches theoretically and experimentally, applied to the adaptation of existing models. The results show that the techniques we present are better for reducing the overall cost than the classical approaches (e.g. oversampling) and show that cost contextualisation can be performed with good results when no data is available. 1.