Setting the Query Cutoff Score

This functionality is used to find the optimum model score cutoff (the score that separates the two labels) for your application. It does not change the constant threshold score of 0.5 that is used to separate the two labels during model training. The Query Cutoff Score group box shows the adjusted error rates for the validation dataset if only record pairs with the prediction score at or above the cutoff score were classified as matches. The adjusted false positive rate shows the percentage of predicted matches that are not true matches. The adjusted false negative rate shows the percentage of true matches that would be classified as false. The adjusted error rate shows the overall percentage of incorrect results for the current cutoff score.

In most situations it is a best practice that you start using the model with the default cutoff score of 0.5, which reflects actual predictions made by the model and usually minimizes the overall error rate. However, your particular application might require that either false positives or false negative be reduced to an absolute minimum. Using this function you can determine the suitable cutoff score that achieves the required false positive or false negative rate.

The Optimize button sets the cutoff score to minimize the overall error rate. The Default button sets the cutoff score to the training threshold level of 0.5.

Figure 48: Query Cutoff Score