DoorDash introduces an ML model to predict a store’s operational status to increase user experience and save thousands of canceled orders.
Understanding a merchant’s operational status and ability to receive and fulfill orders is critical to the DoorDash platform. Each store or outlet running the platform is self-contained, and information about operating hours is important for the customer to see that their order cannot be delivered and to avoid submitting other orders to closed stores.
How DRSC works
When dashers find that they cannot get an order to a store location that appears to be closed, they are prompted to take a picture of the store to start the reporting process. When a valid photo is uploaded, the dasher is compensated for the partial time and effort they spent to get to the store and is signed out of the delivery so that they can continue their dasher and be assigned other deliveries.
When a DRSC report is received, a number of actions can be taken in respect of the order: the delivery can be cancelled and the customer reimbursed.
In parallel, Platformtra contacts the trader to confirm that the shop is indeed closed. If the merchant confirms the closure or does not respond, the store is removed from the app for a certain period of time.
Since some DRSC reports are inaccurate because the image does not show a closed store or the merchant confirms that it is in fact open, having a reliable means of reviewing the validity of the DRSC means that we need to better prepare to assign another Dasher.
Because some DRSC reports are inaccurate because the image does not show a closed store or the merchant confirms it is actually open, having a reliable means of reviewing the validity of the DRSC meant we would be better prepared to assign another Dasher and complete the order when the validity of the DRSC report was in question.
Many companies have built their operations based on heuristic methods or simple rules. Rules are the most direct path to proving that a problem is solvable by going from nothing to functional implementation. However, there can be a large opportunity cost if one does not consider upgrading the heuristic solution into an optimized ML solution that will increase performance to a near-maximal level.