From Talk to Tangible. A Real-World Guide to Machine Learning.
Part 5: Execution
In the last post, Engagement, I shared two interrelated approaches to validating your Machine Learning strategy, i.e. Proof of Concept to ensure the data science is possible and Proof of Value which demonstrates real-world results. Execution will review how you promote the Machine Learning capability into production and integrate it with your business processes.
Getting the Right Insights to the Right Place at the Right Time and in the Right Form
Congrats, you’ve made it — you designed, developed and validated the model.
Although, the days of data science teams taking months to produce analysis, then handing over a static spreadsheet or PDF are over. Today, we must cover a few areas to ensure insights are incorporated into the business. First is automation, sometimes referred to as the analytics factory, where data is automatically ingested, formatted, and fed into Machine Learning for execution, then the output is automatically distributed to the appropriate destination.
The second component is form and timing. In this case it is critical to deliver insightful information how and when the user needs it delivered. This means if the the user lives in their email inbox, then you should send the insights via email. However, if the user has a home system, e.g., care-management system, then you should learn about the potential APIs for direct app-delivery. Other times Machine Learning’s score or prediction may feed into another system that includes the insight into additional calculations such as lead scoring apps.
The third and final point is to ensure there is a mechanism in place to measure the impact of the Machine Learning capability on the business. To grow your Machine Learning capabilities over time, you must continually demonstrate measurable before and after evidence. Example, customer Life-Time Value was $10,000 before implementing the LTV model, and over the last three months LTV is now $11,500, or a combined $1.1 million, top line impact.
Overview of an Analytics Factory
The specific technical approach and decisions are best left to your data science team, but you should be aware of the key components of an analytics factory.
Machine Learning thrives on variable and diverse data, which means the factory needs to be able to ingest and process all types of data — structured data-like tables, unstructured data-like SMS text as well as images and voices files. Landing data in a reliable and scalable storage environment such as AWS S3 is essential because it enables your team to focus on more strategic efforts.
Landing data is not the same enriching data, some use AWS Redshift or Azure SQL to create robust analytics storage environments where the team can join multiple types as part of the data pipeline. As mentioned above, different users will have different needs, so the factory needs to be able to produce compelling data visualizations with tools such as Tableau or D3, or simply email spreadsheets at specific times.
There are numerous Machine Learning platforms and frameworks — everything from open source libraries such as Scikit-learn to Machine Learning frameworks, such as MXNet and TensorFlow, as well as managed platforms such as AWS SageMaker — no matter your needs there are multiple options.
Where we can go from here
Once you have your model in production, you will want to improve the model’s computational performance and accuracy as well as user adoption — in the next post we will dive into Enhancing, our 6th and final E.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — -
David is the practice lead for Advanced Analytics and Data Visualization for Slalom Philadelphia. Slalom is a modern consulting firm focused on strategy, technology, and business transformation.We help companies tackle their most ambitious projects and build new capabilities. @slalomphilly
