Also referred to as model serving, model scoring, putting the model into production, model operationalization, model integration, model implementation, model productionalization, or predicting, model deployment is the process of integrating a machine learning model into a business process to make practical business decisions based on new data, e.g., identifying clients to make retention offers to or scoring the likelihood of a machine part to fail. Deployment can be as simple as producing an Excel report or complicated as automating an elaborate business process.
Model deployment comprises putting model code into a format compatible with existing business systems and building the required data and monitoring infrastructure. Direct coding into a required format can take weeks, even months, to put a model into production because it often entails re-coding the entire model from the language it was written in (e.g., R, Python, Scala, etc) to the languages the enterprise production system can understand (e.g., C++, Java, etc). However, using automatic machine learning (AutoML) platforms such as RapidMiner, KNIME, DataRobot, H2O.ai, Anaconda Enterprise, and Dataiku, can take a few minutes to put a model into production and monitoring.
Goals #
- Integrate the model into business workflow
- Ensure stable model performance over its deployment period
Tasks #
- Implement trial deployments and experiments to ascertain deployment stability and model performance before switching to new model
- Deploy model in the way that best satisfies the business needs of the project, e.g., as a file, as a input to external application, as a dashboard, etc
- Automate as much of the deployment process as possible to ensure reproducibility and reliability
- Collect output of deployed model and feed it back into the training process
- Track quality of model predictions, i.e, model drift, substandard throughput, sudden changes in performance, etc
- Draw insights from model outputs and performance and correct faulty assumptions, conceive new hypothesis and/or inform improvements in model architecture
- Monitor model for safety and compliance with legal and regulatory requirements, i.e., access control, model lineage, audit trails
Roles #
- Project Sponsor
- Project Manager
- Data Engineer
- Machine Learning Engineer / Data Scientist
- Business User
- Domain Expert
