This is a guest post from Jodi Blomberg, a Principal Technical Architect at SAS. She has over 12 years of experience in data mining and mathematical modeling, and has developed analytic models for many government agencies including child support enforcement, insurance fraud, intelligence led policing, supply chain logistics and adverse vaccination reactions. Enjoy her practical advice! - Paula

At the INFORMS Conference on Business Analytics and Operation Research in Chicago I gave a talk called “Starting at the End: Letting Results Drive the Analytic Process". It focused on how to be “results oriented” in your approach to data mining.

For all of the excitement and buzz about analytics, analytics projects (not just in government) have a less than impressive success rate. This is rarely due to the lack of good models or good analytics. It is more often about a failure to implement the results of the analytic models- even when the models themselves are statistically valid and contain good actionable information.

For nearly every analytics project, the goal is the same: to provide information that will change the way someone does their job or makes decisions related to their job. While the average data mining analyst or project lead cannot always insure that the organization is committed to changing the workplace, they can do better at setting up their model for a successful implementation.

Often, projects with good actionable results fail or just fail to be used because the results were not clearly defined before starting the project. If no one knows what to expect, they are less likely to use the analysis. Tweaking the traditional analytic process to involve stakeholders and to increase the likelihood that the analytic models will be implemented is a good start. Getting folks to think about what happens after the results are in -- what to do with them or who might do it -- is key, and it’s better done earlier than later.

In my talk I discussed success and how to measure it, the first steps to achieving it. I gave examples, including the highlights of my favorite mistakes. Often, modeling mistakes can be caused if you forget about the "people effects".

For instance, I worked with a customer on analyzing where to place ambulances to ensure the entire population had coverage. The initial solution, which focused only on optimal ambulance placement, would have ambulance drivers constantly repositioning their rig. This would preclude them from ever engaging in activities like eating lunch and going to the bathroom. Not a very practical solution! To fix this, we restricted our solution to attempt to find an optimal placement of ambulances such that if an ambulance went out on the call, we would attempt to find an optimal repositioning that moved only one ambulance. This way, most of the rigs could stay where they were when not on calls.

While a mistake, it's cases like that where analytics collides with practical reality that can be the most educational and fascinating.