More and more, organizations are relying on machine learning (ML) models to turn massive volumes of data into fresh insights and information. These ML models are not limited by the number of data dimensions they can effectively access and use vast amounts of unstructured data to identify patterns for predictive purposes.
But model development and deployment are difficult. Only about 50% of models are ever put in production and those that are taken at least three months to be ready for deployment. This time and effort equal a real operational cost and mean a slower time to value, too.
All models degrade, and if they are not given regular attention, performance suffers. Models are like cars: To ensure quality performance, you need to perform regular maintenance. Model performance depends not just on model construction, but also on data, fine-tuning, regular updates, and retraining.
Here ModelOps come into play.
According to Stu Bailey, Co-Founder and Chief AI Architect of ModelOp, “ModelOps is a capability that focuses on getting models into 24/7 production. It’s a capability that must be owned by the CIO’s organization or the technology center of a large organization.”
ModelOps allows you to move models from the lab to validation, testing, and production as quickly as possible while ensuring quality results. It enables you to manage and scale models to meet demand and continuously monitor them to spot and fix early signs of degradation. ModelOps is based on long-standing DevOps principles. It’s a must-have for implementing scalable predictive analytics. But let’s be clear: Model development practices are not the same as software engineering best practices. The difference should become clearer before its implementation.
Need to Understand Underlying Terms for ModelOps Implementation
For a ModelOps first step, you need to monitor the performance of your ModelOps program because ModelOps represents a cycle of development, testing, deployment, and monitoring, but it can only be effective if it’s making progress toward the goal of providing the scale and accuracy that your organization requires.
You need, at the highest level, to determine the effectiveness of your ModelOps program. Has the implementation of ModelOps practices helped you achieve the scale, accuracy, and process rigor the organization needs?
Then, at the operational level, you need to monitor the performance of each model. As they degrade, they will need retraining and redeployment. Here are some considerations when creating a performance dashboard:
• For models (or classes of models), set accuracy targets and track them through development, validation, and deployment for dimensions such as drift and degradation.
• Identify business metrics affected by the model in operation. For example, is a model designed to increase subscribers having a positive effect on subscription rates?
• Track metrics such as data size and frequency of updates, locations, categories, and types. Sometimes model performance problems are due to changes in the data and its sources and these metrics can help in your investigation.
Monitor how much computing resources or memory models consume.
Related to metrics, model validation is an important foundation of ModelOps. Some use validation and verification interchangeably, but their intent is different.
*Note: Verification confirms that a model was correctly implemented and works as designed. Validation ensures the model provides the results it should, based on the model’s fundamental objectives. Both are important best practices in the development and deployment of quality models.
Problems to Address While Using ModelOps
As noted by SAS, models can degrade as soon as they are deployed (sometimes in days). Of course, some things will affect the performance of your models more than others. Below are some common problems – ones that you will almost certainly encounter.
Subtle changes or shifts in data that might go unnoticed or may have a lesser effect on some traditional analytical processes can have a more significant effect on machine learning model accuracy.
As part of your ModelOps efforts, it’s important to properly assess the data sources and variables available for use by your models so you can answer:
• What data sources will you use?
• Would you be comfortable telling a customer that a decision was made based on that data?
• Do the data inputs directly or indirectly violate any regulations?
• How have you addressed model bias?
• How frequently are new data fields added or changed?
• Can you replicate your feature engineering in production?
Time to Deployment
Because the model development/deployment cycle can be long, you first assess how long that cycle is for your organization, then set benchmarks to measure improvement. Break down your process into discrete steps, then measure and compare projects to identify best and worst practices. Also, consider model management software that can help automate some activities.
Be on the lookout for things like drift and bias. The answer to these problems is creating a strong approach to model stewardship in your organization. If everyone from model developers to business users takes ownership for the health of your models, these problems can be addressed before they affect your bottom line.