When training a model that contains a Boolean goal variable, the system uses an iterative process to select a true/false threshold that will generate the most accurate model. That is a model that finds the most true positive and true negative results.

However, in some scenarios, a straight forward classification of the data does not yield the most useful model. By customizing the confusion matrix weights, you can optimize the selection of the Boolean threshold in a way that generates a model more suited to a given business scenario. Consider the scenarios below.

Suppose you are trying to identify patients with a specific diagnosis from large dataset. The number of patients who actually have the disease may be quite low, but the consequences for missing the diagnosis in a single patient could be serious. By setting a higher True Positive weight during model training, you can influence where the Boolean threshold is set to optimize the classification of true positive cases.

Suppose you are trying to predict a machine failure. You estimate that the cost of not predicting a failure (false negative) is about $90 in unplanned downtime and service charges. The cost for incorrectly predicting a failure (false positive) is a $20 service charge. You can customize the matrix weights to act as a cost matrix that will optimize the model based on the costs of false positive or false negative results.