The application of machine learning in computing has been one of the fastest-growing areas of information technology. Experienced government IT leaders understand that mining insights from data can transform their agency. I've experienced this same understanding as a professor at the University of Minnesota, when I tried to make machine learning work on enterprise applications. I’ve since learned that machine learning is not an extreme sport.
Here are ten ways to set up your government agency for success:
1.Build a data lake and infrastructure with a clear goal.
Bigger projects take longer and have a higher risk of failure. A strategy to "boil the ocean" and aggregate every single piece of enterprise data into one place is a recipe for failure. Recalibrate and focus on a small set of data and a narrow problem: aggregate, understand, and analyze that data first, then apply that understanding as you scale your efforts.
2. Don’t get distracted by shiny new hardware/software tools.
Writing lots of code isn't agile. Getting results quickly is. In machine learning, the gulf between vendors that can help you deliver results quickly and efficiently and vendors that can't is enormous. Concentrate on tools that get results.
This same principle applies to hardware. At one time, the NoSQL and big data "movement" espoused that building a Hadoop data lake and running NoSQL tools would yield impactful results quickly. That simply wasn’t true.
3. Embrace the value of SQL and relational databases
Historical note: The NoSQL industry once suggested that SQL couldn't be used at scale to do machine learning. Yet every major cloud vendor acknowledged that SQL not only could be scaled, but provided transactions, reporting, clean data practices, and most importantly, simplified programming that NoSQL products never could.
SQL is where most of your data already lives. Consider products like Oracle's database and Exadata platform that allow you to run machine learning directly on your data leveraging SQL and other high-productivity tools, without the need to build separate infrastructure, tools, and ETL processes.
4. Ensure your ML data is secured, privacy-protected and enterprise-governed
Moving data out of secure enterprise databases to insecure "big data" lakes is so risky most enterprises refuse to do it. Keeping your enterprise data in a secure data like Oracle's, where you can perform machine learning in-situ, removes the need for the complexity and overhead of copying while delivering machine learning results to the business quickly and safely.
5. Start with a small, focused business problem
Getting insights from data is hard. Focus on a small problem with an immediate business impact. Show the power and potential of machine learning, without burning piles of dollar bills in the process. Learn more from these state and local government customers.
6. Build only when necessary
Reuse code whenever possible. Buy as much as possible, only build when necessary. If you think that writing lots of code to get your machine learning results is the right strategy, the business is going to be disappointed and you are at risk of huge cost overruns and delays.
7. Watch for strong signals in the data for KPIs you care about
Make sure you get the right data that has the signals you need to optimize for business results. Machine learning can't work without the right data. Run small experiments and use common sense to find the right input data for your problem. Experienced data scientists can add a lot of value in helping you find signals.
8. Enable citizen data scientists
Don't assume you must have a priesthood of data scientists to find impactful patterns in your data. Your line of business analysts have the best insight into what these patterns might be, and automating the model-building process for these teams mean you get to business-oriented results faster, without the pain, cost, overhead and delays of a an extended science project.
9. Leverage automation to build your ML models
Tools like Oracle's Auto-ML allow your business analysts to build machine learning into your applications without hiring lots of hard-to-find data scientists.
10. Plan to tie machine learning results back into the business
Even when businesses find models that work, making them operational can be very difficult. Leverage tools that allow you to inject any machine learning insights you gain immediately back into operational work flows, in real time.