Embedding Inductively Developed Models in Transactional Database Systems: Examples Using Multi-Strategy Classification Models

摘要

Multiple frameworks have been developed for the comparison of classification models that account for the overall accuracy of decisions made by models. Among such frameworks are confusion matrices and ROC (receiver operating characteristic) curves. These frameworks adequately capture trade-offs between a model's specificity and sensitivity. However, as inductively developed classification models are increasingly working their way into transactional environments, an additional element requires consideration, namely performance, or the speed with which accurate decisions can be made. This paper evaluates alternatives for the deployment of a neural network in a transactional environment, thus establishing the need to explore performance as a constraint to more traditional methods of model assessment.