Overview

Since Zementis' (www.zementis.com) solutions provide a painless and rapid method of deploying predictive analytic and machine learning models I occasionally get asked about the final phase of CRISP - DM. CRISP-DM is the Cross Industry Standard Process for Data Mining. The final phase is deployment of predictive analytics and/or machine learning models.

Below is a checklist I have developed to assist organizations in evaluating how they deploy advanced analytics.

1. Confirm the production information technology (IT) environment where analytic models will be deployed.

2. Confirm who will be responsible for deploying and maintaining the analytic models in the operational / production environment.

3. Export model to organization’s standard format if there is one. Consider using the open industry standard Predictive Model Markup Language (PMML) managed by the Data Mining Group (DMG). More information at www.dmg.org.

4.   If not using standard format investigate whether or not the model as written in the development environment will run effectively in the operational / production IT environment.

5.   If the model will run as written provide to the responsible party for   deployment. Ensure communications are complete to the consumer of the model results as to when the new model will be operational.

6.  If the model will not run as written contact the correct liaison in the IT organization and begin the process of working with them to custom-code the model to run effectively in the operational environment. Another option is to evaluate systems that enable the immediate deployment of models without custom-coding into operational / production IT environments such as Zementis solutions, yes this is shameless self promotion.

7. Criteria for evaluating model performance over time should be determined.   The party responsible for monitoring the model and how often should be agreed upon within the organization prior to deployment.

8. The likely timeframe for revision for the model should be determined at time of model deployment. Data Science and IT Human resource allocation should be budgeted for at time of deployment.  This becomes critical as the number of models under management increase and is why companies looking to increase both competitiveness and data science productivity bring on efficient and effective deployment solutions for machine learning and predictive analytic models.

Keep in mind if custom coding, it is often helpful for the developer to understand and/or have documentation relating to CRISP_DM phases I – V.