Angelov, Boyan. (2020, December 8). Shotgun MVP: Reducing risk in highly unpredictable ML products (Version 1.0). Zenodo. http://doi.org/10.5281/zenodo.4311945
Machine learning products do not fit neatly into the agile methodology. There are different reasons for this, the primary one being the lack of certainty behind which directions will be successful (due to data quality and similar issues). Thus, what often happens in organizations is pilotitis - the disease of launching MVP after MVP without tangible progress and commitment.
There is a method to address this issue while still alleviating the uncertainty behind planning MVPs - the Shotgun MVP. As a first step, several MVP initiatives are launched simultaneously (staffed to the bare minimum needed). Thus, all possible approaches are covered. This process continues for several sprints until a clear frontrunner becomes apparent - based on a “success” metric. There is flexibility in defining this metric, and possible examples include performance metrics (i.e., model accuracy) and solution complexity (such as the workforce needed to complete further development iterations). After this MVP (MVP II in our case) is selected, resources are committed fully to it for the following sprints while keeping the other options as backup plans or possible enhancements for downstream work.
This simple procedure is not easy to perform and requires careful and restrained execution but can provide the shortest way to delivering a successful ML product.