Recent years have seen the burgeoning growth of data science and ‘insights’ teams to make sense of the large amounts of data that corporations have at their fingertips.
This is often a costly exercise; corporations are funding teams of highly specialised professionals to do a job the business does not entirely understand, working with data that was never collected for business intelligence purposes. In addition, “machine learning” seems to be latest cure all catch cry, where management believes that computers will begin teaching “consumer 101”.
However attempts to utilise data often lead to disappointment; it can be hard to justify the costs of developing and implementing an analytics strategy on what seems like a gamble, and many who utilise analytics do not feel they are getting what they want out of their data science teams. The problem lies in the disconnect between analytics and business objectives.
Machine Learning is great, but is not always the solution
Machine learning has its place, but often traditional methods can be utilised instead and achieve approximately the same accuracy. The key advantage of using traditional statistical methods is that it becomes much easier to explain to stakeholders how it works without going into a maths lesson. Remember that the end goal is to provide information to make better business decisions, not necessarily to predict a response variable with the highest accuracy.
In addition, the results often highlighted in the media only apply to specific examples. As such, a statistical method can often be a better predictor than a machine learning method. Take linear regression, a technique learnt by most business and science students in a mandatory introductory stats course, versus random forests, a common machine learning technique. While there are many cases in which a random forest will outperform regression, it should be recognised that just like traditional statistics, machine learning has limitations.
If your data set does not predict your response variable, it does not matter what analysis techniques you use. A large amount of weight is placed on the latest and greatest methods, and not on the wider design of the analytics system. You are substantially better off designing a new system to collect data with strong predictability, than attempting to build an analytics system that works on a legacy database that focused on collecting information solely for operational purposes (usually only basic demographics and some interaction data).
Solutions should make sense to experienced stakeholders; any “assumption busting” outcome should have a clear and reasonable explanation as to why the results are different to what was expected. Industry experts are unlikely to adopt analytics unless they make sense at a conceptual level. The world of consultancy likes to obscure methodology by “protecting company IP”, which results in solutions being a “black box” that no one understands. This is often counterintuitive, as solutions are left unused and often regarded a waste of money. A simple solution is not necessarily inferior, particularly if you can get widespread company buy in.
You only need to beat a coin toss
It may sound dull, but even models with weak prediction capability can still make a huge difference in business activity. While it is not ideal, many businesses see poor data quality and the lack of a guaranteed outcome as barriers to using analytics to inform business decisions. However, even simple analytical models can transform business decisions by highlighting different consumer behaviours for product development, reduce marketing campaign lists, and identify which customers are most likely to churn.
Analytics and machine learning are great tools to better inform business decisions, but it is important to not get caught up in all the hype. No, you can’t outsource your job to a computer, nor trust every result that it provides without additional research and context. At CoreData, we often get asked how to include analytics in a project without clearly defined business objectives. In the “information age” we have an abundance of solutions, but often none that solve our problems. The answer? Start with the problem, not the solution.