The Model Management page of IBM® SPSS® Collaboration and Deployment Services Deployment Portal provides the ability to monitor the ongoing performance of models deployed to IBM SPSS Collaboration and Deployment Services Repository. These model files are associated with jobs that can be executed on demand or scheduled. The files are created by using IBM SPSS Modeler. Model evaluation and champion challenger jobs are set up and executed using IBM SPSS Deployment Manager, and IBM SPSS Collaboration and Deployment Services Deployment Portal is used only to view the results. The information displayed as panels on the Model Management page can include the following:
- a listing of the best and worst performing models
- trends of model performance
- champion models
- a listing of all available model files
The options on the Configure panel can be used to specify display parameters and show or hide individual tabs.
Model evaluation and comparison can focus on accuracy, gains, or accreditation.
- Accuracy. The accuracy of a model reflects the percentage of target responses that are predicted correctly. Models having a high percentage of correct predictions are preferred to those having a low percentage.
- Gains. The gains statistic is an indicator of the performance of a model. This measure compares the results from a model to the results obtained without using a model. The improvement in the results when using the model is referred to as the gains. When comparing two models, the model having the higher gains value at a specified percentile is preferred.
- Accreditation. Model accreditation reflects the credibility of a model. This approach examines the similarity between new data and the training data on which a model is based. Accreditation values vary from 0 to 1, with high values indicating greater similarity between the predictors in the two data sets. When comparing two models, the model having the higher accreditation value is based on training data that is more similar to the new data, making it more credible and preferred.