From Data to Decisions: How ModelOps Can Get You There Faster
ModelOps can help your enterprise cultivate a new way for data scientists and IT to work together and deliver faster, tangible business outcomes.
- By Bryan Harris
- October 18, 2021
Today's competitive, data-driven environment is demanding that organizations put increasing numbers of analytics models into production more rapidly than ever before. At the same time, data scientists are charged with ensuring trust, repeatability, and transparency in those models. This is where ModelOps, a superset of machine learning operations (MLOps), comes in to help identify, address, and solve organizational complexities and provide a framework for repeatedly moving models into production.
This emerging approach focuses on effective operationalization of all types of AI and decision models, and supports continuous retraining, automated updating, and synchronized development and deployment of more complex machine learning models. The result: greater model productivity and efficiency that moves organizations from data to decisions faster.
Creating a Model for Success
Data scientists work within an environment that automates, governs, and accelerates the analytics life cycle to deliver the greatest value. If an organization is spending the majority of its time stitching together technology and manually performing repetitive tasks, it can lead to lost opportunities for improved productivity and success.
The good news is that ModelOps greatly improves an analytics project's chances for success by bridging the gap between the analytics and production teams. The approach eliminates silos, streamlines handoffs, promotes quality, and, above all, ensures scalability. By moving analytics models from the data science lab through registration and deployment as quickly as possible, ModelOps drives high-quality analytics results, enables better decision making, and improves business outcomes.
Analysts agree. According to a recent report from analyst firm Forrester, the rapid maturation of MLOps tools is leading to a "breakout year" for ModelOps. As Gartner researchers put it in an August 2020 report ("Innovation Insight for ModelOps"): "ModelOps lies at the center of any organization's enterprise AI strategy."
Embracing the Benefits
How organizations use analytics models changes and matures over time, but the process remains essentially the same. Most begin with a few data analysts. Leadership starts to pay attention to the analytics results; data starts to be organized, better technology is introduced, and a few models are built. When this investment generates business value, demand for faster analytics increases and pressure for new models grows.
At this point, how quickly data scientists can identify the right data, explore the ideal algorithms, and ultimately train one or more models is paramount. The longer this entire process takes, the more it delays the business impact as well as increases the risk that models in production are working on assumptions in the data that are no longer valid.
On the plus side, although ModelOps requires a departure from business as usual, in today's data-driven digital economy, no other endeavor will be faster at deploying more effective analytics models and delivering business value. Specifically, benefits include:
- Increasing transparency of decisions with responsible AI
- Improving productivity through automation
- Accelerating insights with high-performance architecture
- Reducing risk with end-to-end governance
An excellent business use case for ModelOps is USG Corporation, a North American producer of gypsum wallboard, joint compound, and many more products for the construction and remodeling industries. The manufacturer deploys optimization models to analyze plant inputs (such as flow rates and raw material additives) to predict quality outcomes before production even starts. Using ModelOps, USG can test the performance of challenger models, select the top performer, single out the optimal formulation of raw materials, and adjust its production process in near real time. As a result, USG manufactures products that meet its quality standards at the lowest possible price.
Getting Started with ModelOps
If you are ready to get started with implementing ModelOps at your organization, consider these best practices:
- Think of deployment as you begin development. The target operational state should not be an afterthought. Build a robust model management and monitoring capability as part of ModelOps.
- Connect model metrics to business KPIs. The purpose is to generate business value, not build science experiments.
- Leverage a platform that supports an end-to-end analytics life cycle. ModelOps isn't just about retraining existing models. A solid ModelOps process enables the consideration of new models, new analytics techniques, and even new data sets.
- Strive for consistency, accuracy, and trust in your decisions. An overarching goal of ModelOps is to scale the deployment of models throughout the organization. Doing this consistently and accurately will ultimately build confidence and trust in the value of machine learning and AI. Simply put, businesses need to know they can trust the models their data science teams are creating.
- Establish an effective process to ensure governance and transparency. Organizational trust is foundational to the analytics process and dependent upon solid data, model governance, quality processes, and an ability to foster innovation among users. A lack of governance, documentation, and interpretability makes it much harder to know if models are performing well.
Like all good things, you can't implement ModelOps overnight. Adopting this approach takes time and presents unique challenges to each organization, depending on its size, existing systems and processes, and culture. However, with each step, ModelOps will cultivate a new way for data scientists and IT to work together and evolve an analytics program to deliver faster, tangible business outcomes.
Bryan Harris vice president and chief technology officer at SAS. Harris has more than 20 years of experience researching and developing analytics techniques, enterprise search technologies, distributed computing and cloud architectures, and user experiences for both the federal and commercial industries. For nearly 10 years, he has been a critical senior leader of SAS R&D. You can reach Bryan on LinkedIn and Twitter.