Donatos was sitting on a wealth of customer data, including demographic information, what they ordered, how they paid, time of order, time promised, cost of purchase, complaints, and much more. This abundance of data made it easier for Donatos and our Fusion team to explore a machine learning model in selected stores across the country. The pilot program also included a control group for comparison purposes.
Creating and implementing a machine learning model involved:
- Putting Donatos’ extensive data on a cloud platform that would accelerate the process
- Loading and landing the data to let the machine learning algorithms do their job
- Evaluating and selecting the Donatos data most capable of providing accurate answers. (This step included pulling in the source data, aggregating it, and then filtering it for aberrations, such as orders not expected to repeat, like an out-of-town business person.)
- Assessing the quality and quantity of the data
- Cleansing the data to use as a training set, which was used to identify which machine learning algorithm would produce the most accurate model to predict who would stay or leave
With the foundation set, each day we’d run the previous day’s sales in each of the pilot stores against this model to produce a list of customers who were highly likely to leave. Store managers then took action to get these identified customers to return. Though it was a short trial, the results were impressive.