By: Kelley Counts, OneBlood
There is a hard lesson many data scientists are learning when they start their first job – and no it’s not because they lack ninja skills. While there may be more than a few data scientists that could use some pointers on self-defense, I am not suggesting martial arts is completely necessary to be a true data ninja. But that would be pretty cool. I’m talking about the ability to leverage analytics solutions to solve real world problems for organizations. AND – The catch is you have to do it while maintaining an analytical pace the organization can feel comfortable keeping up with.
As the title proclaims – Bruce Lee could certainly teach us a few things about that. Without a doubt, he is noted as an influential movie star and martial artists, but he was also just as accomplished at philosophy. I’ll get to the point by referencing one of his quotes –
“Empty your mind, be formless, shapeless like water. You put water into a cup, it becomes the cup. You put water into a bottle, it becomes the bottle. You put it into a teapot, it becomes the teapot. Now water can flow or it can crash. Be water, my friend.” Bruce Lee.
Your analytics solution should be water and the organization the cup. By no means should it be drastically the other way around. Especially if the organization is new to advanced analytics. The more you can make water become the cup the easier it is for the organization to get on board with higher adoption and less resistance.
Analytics solutions are excellent at creating value and reducing inefficacies. But, if they require the organization to make drastic changes they are not ready for or do not completely understand, they can be excellent at creating inefficacies and reducing overall business value. As Bruce Lee said, water can flow or it can crash.
The analytical solution should fit the organization, just like water becomes the cup so it can flow to create business value, if it doesn’t, it just might crash.
How To Flow & Not Crash.
1. Put the analytics solution just a little bit ahead of the organization. But not too far ahead.
Everything seems more complicated these days, that’s because it is. I’m not necessary talking about the complexities of analytics, but rather organizational change. Truth be told – you should understand how an organization with embrace analytics solutions just as much as you understand the solution. If the solution causes operational change, as it usually does, then make organizational adoption a high priority. Think small steps in a big direction that stakeholders understand, are interested in, and can learn from. If a data driven culture isn’t alive and well, you will have to foster it. Becoming too sophisticated too fast can have negative effects even if you have the best of intentions. Small consistent wins done well over the long term will build momentum and lead to big results that will keep the organization’s analytics maturity growing at the right pace.
For example, if the organization is not ready to let the black box drive, utilize decision trees so executives can see the logic at which AI is making decisions. You can even do this in a non-deployable fashion by presenting a decision tree chart with the threshold breaks to give a high-level overview for executive consideration. Once a comfort level is developed with the logic, the next step would be a deployment pitch that helps your team understand the importance of accuracy and variables by using the model. Once that understanding is developed you will be on your way to having their trust in a more sophisticated solution as long as you lead in an educational manner.
2. Sometimes less sophisticated solutions that are easy to understand and require minimum operational change are better than sophisticated solutions that are confusing and require greater operational change.You have skills, we get it. Probably not Bruce Lee skills but data skills at the Bruce Lee level – RIGHT. Just like an expert martial artist instructor would not start a new student off with advance techniques because they could crash is the same reason data scientists should not start off organizations new to analytics with extremely advanced techniques. Especially if the techniques are unclear and require a lot of organizational change in the way of operations. Change can be scary, especially if you do not fully understand all the elements around it. This is the same for companies. So, if you have an analytics solution that provides business value – ask yourself (1) is the solution easy to understand, not for you, the company. Then ask yourself (2) does it require significant operational change that could create resistance. If the answers are No & Yes, look into finding a solution that gives you Yes & No answers to start with.
Here is a scenario, if a marketing department is open to hearing your propensity modeling pitch but still not exactly sure how predictive modeling will streamline into their operations, try these approaches. Find something that is being somewhat neglected and has limited barriers in terms of deployment. In other words, operate in an area where you will not annoy anyone if you mess up. For example, if marketing is segmenting customers on a regular basis and has a lot of strategy built into this operationally, something like, daily call lists for equipment sales, that may not be the best place to start. Instead look for something that has a broader horizon such as current customers that could be interested in equipment upgrades within 90 days, especially if this area is being overlooked or not even thought of. This is a good “needle in the haystack” problem, which typically are not as operationally taxing, from an organization perspective to implement predictive analytics on.
You can build the operations around your idea by scoring customers every 90 days and have marketing make an “upgrade” campaign that spans 90 days. In other words, you found a way “in” with low level deployment that is operationally sensitive. Remember, if this is the organization’s first interaction with modeling, do everything you can to keep it simple without sacrificing accuracy. Once everything is humming and you have built departmental trust there will be a higher tolerance level, room, and opportunity for more sophisticated models within a more operationally challenging environment.
3. Start with the business question. It is better to be kind of right about a great business question than exactly right about a poor business question. Be a black belt at bringing business value that causes shifts.
You can ask a lot of business questions which analytics can provide answers too. Focus on the value the answer will bring first, and look at the accuracy of the analytic solution second. Quite often, if the “kind of right” analytics answer brings a lot of value to the organization, then more than likely it is something they need to learn more about to unlock more value. Being exactly right about a poor or easy business question usually gets the “we already knew that” response from executives. When you can create great value with low accuracy don’t get discouraged – get excited, because you may have just found a competitive advantage that can change the industry.
In one particular case, we were struggling to reach a respectable accuracy for a classification model that was going to be used to segment an extremely valuable customer segment. But at the end of the day we were “kind of right” about a great business question. Our direction was to deploy the model using the predictive variables we had. Then investigate feature importance, and focus our attention on breaking down the best variables into small subsets. By staying the course, we were able to “datafy” and “evolve” new predictive variables from the original ones we started with and incrementally increase accuracy over time. In fact, there was a lot of support from operations to find better variables that otherwise would never have been thought of. The special sauce to getting organizational buy in was (1) communication that variables did not completely exist to provide a good answer and that means a big opportunity to create a competitive advantage and (2) by committing to the predictive variable quest we will learn a lot and develop a proprietary data set that will be far ahead of the curve. Let’s face it sometimes CRISP DM does not always fit within the tolerance of an organization’s pace but if the business value is there it sure does make the momentum of the iterative portion much easier.
It is very much an art creating analytical solutions that allow for natural organizational growth and advancement. They have to fit the organization like water fits a cup, and be of such that an organization can reasonably consume and understand at the right pace and tempo. If the solution is too complicated and too far ahead of the organization’s comfort zone, there is a risk that it will not be embraced. When done just right and in sync with the organization there are tremendous growth opportunities.
About the Author:
Kelley Counts is a Data Scientist with OneBlood. He has a Bachelors in Biochemistry, Masters of Science in Forensic Genetics, and a MBA in Business Analytics. He is a contributor and designer for graduate level Predictive Analytics course work utilizing Python. Kelley has frequently lectured around the country on topics including the practical applications of machine learning within a variety of fields. He also serves on international professional boards as an advisor for matters relating to digital technology, information systems, and analytics. His current work is focused on pioneering innovative machine learning applications in the pharmaceutical transplant fluid industry.