Here are some points that I have shared with my Prescriptive Analytics students at Bucknell University based on a piece by Gurobi Optimization CEO Ed Rothberg for Forbes:
1) Many applications of machine learning tend to be consumer-facing, whereas mathematical optimization is usually applied in businesses to seemingly optimize their processes without consumers being aware of it.
2) Mathematical optimization can be used to develop a “digital twin” of some organizational processes in situations where the number of possible solutions could be very large and subject to changes, especially in situations where patterns from historical data could be disrupted while the process in the organization remains the same.
3) The initial time and effort required from stakeholders to build a mathematical optimization model can be greater than that of a machine learning model, since these models require a deeper understanding of the business processes involved.
4) Machine learning can leverage what is called “big data” to learn from the past and make predictions about the future, but it is vulnerable to model drift: the predictions become less accurate if, for example, there are changes in the patterns that were previously observed in the data.