Granted Patent B2

US 12,254,482 · App. 18/159,249 · Granted Mar 18, 2025

Inventors: Jacob Solotaroff (Palo Alto, CA); Jamie Rapperport (Palo Alto, CA)

Assignee: Maplebear Inc.

G06Q30/0206G06Q30/0211G06Q30/0255G06Q30/0271

Abstract

Claims (70)

1. A computer implemented method comprising:

receiving, at an offer generation system, data describing an offer on a product;

applying a natural language processing (NLP) model to the received data to extract the product and the offer;

accessing a plurality of test offers stored by an offer bank;

accessing transaction logs related to the product;

scoring each test offer in the offer bank against the extracted product, wherein scoring a test offer of the plurality of test offers comprises:

predicting a forecast score for each test offer by:

applying a reinforcement learning model to the transaction logs related to the product, wherein the reinforcement learning model is trained using transaction logs associated with the plurality of test offers to predict a likelihood that a test offer will achieve an offer objective;

determining a difference between the offer and the test offer; and

updating the forecast score by applying a penalty to the forecast score based on the determined difference;

selecting a subset of the plurality of test offers based in part on the forecast score of each test offer, wherein assigning the subset of test offers comprises maximizing orthogonality of a set of variables associated with the product;

transmitting the subset of test offers to client devices of users;

receiving, from client devices of the users, responses to the subset of test offers, the responses comprising at least one of (1) whether or not a user viewed a test offer, (2) whether or not a user clicked through a test offer, (3) whether or not a user deleted a test offer, (4) whether or not the user saved the test offer to an electronic coupon folder, (5) whether or not a user forwarded the test offer to someone else, (6) whether or not a user posted a test offer to an online forum, or (7) whether or not the user purchased a product using the test offer;

storing the received responses in a database; and

retraining the reinforcement learning model based on the responses stored in the database, such that the reinforcement learning model automatically learns from user responses and continuously improves effectiveness of selection of test offers.

2. The method of claim 1 , wherein assigning the subset of test offers to a plurality of retail locations comprises deploying the offer and the selected subset of test offers in a plurality of retail locations.

3. The method of claim 2 , wherein the deploying is performed to maximize orthogonality between the following variables: store sales, store out of stock rates, number of relevant SKUs carried in each store, temporal effects, discount depth, buy quantity and offer structure.

4. The method of claim 1 , wherein the reinforcement learning model uses Thompson sampling.

5. The method of claim 1 , wherein predicting a forecast score for a test offer by applying a reinforcement learning model to transaction logs comprises:

adjusting the transaction logs for compliance by a retail location in a plurality of retail locations, estimated out of stock events, normalized across stores to account for different store attributes, and adjusted for temporal effects.

6. The method of claim 5 , further comprising wherein predicting the forecast score for a test offer comprises:

determining a lift and standard deviation for the test offer based on the reinforcement learning model.

7. The method of claim 6 , wherein the forecast scores are a baseline function of time from the transaction log data plus elasticity from cross store experiments times a change in price, where in the elasticity is calculated as a function of the lift, and a confidence for the forecast score is calculated as a function of the standard deviation.

8. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the steps of:

receiving, at an offer generation system, data describing an offer on a product;

applying a natural language processing (NLP) model to the received data to extract the product and the offer;

accessing a plurality of test offers stored by an offer bank;

accessing transaction logs related to the product;

scoring each test offer in the offer bank against the extracted data, wherein scoring a test offer of the plurality of test offers comprises:

predicting a forecast score for the test offer by:

applying a reinforcement learning model to transaction logs related to the product, wherein the reinforcement learning model is trained using transaction logs associated with the plurality of test offers to predict a likelihood that a test offer will achieve an offer objective;

determining a difference between the offer and the test offer; and

updating the forecast score by applying a penalty to the forecast score based on the determined difference;

selecting a subset of the plurality of test offers based in part on the forecast score of each test offer, wherein selecting the subset of test offers comprises maximizing an orthogonality of a set of variables associated with the product;

transmitting the subset of test offers assigned to client devices of users;

receiving, from client devices of the users, responses to the subset of test offers, the responses comprising at least one of (1) whether or not a user viewed a test offer, (2) whether or not a user clicked through a test offer, (3) whether or not a user deleted a test offer, (4) whether or not the user saved the test offer to an electronic coupon folder, (5) whether or not a user forwarded the test offer to someone else, (6) whether or not a user posted a test offer to an online forum, or (7) whether or not the user purchased a product using the test offer;

storing the received responses in a database; and

retraining the reinforcement learning model based on the responses stored in the database, such that the reinforcement learning model automatically learns from user responses and continuously improves effectiveness of selection of test offers.

9. The computer-readable medium of claim 8 , wherein the instructions for assigning the subset of test offers to a plurality of retail locations causes the processor to:

deploy the offer and the selected subset of test offers in a plurality of retail locations.

10. The computer-readable medium of claim 9 , wherein the deploying is performed to maximize orthogonality between the following variables: store sales, store out of stock rates, number of relevant SKUs carried in each store, temporal effects, discount depth, buy quantity and offer structure.

11. The computer-readable medium of claim 8 , wherein the reinforcement learning model uses Thompson sampling.

12. The computer-readable medium of claim 8 , wherein the instructions for predicting a forecast score for a test offer by applying a reinforcement learning model to transaction logs further causes the processor to:

adjust the transaction logs for compliance by a retail location in a plurality of retail locations, estimated out of stock events, normalized across stores to account for different store attributes, and adjusted for temporal effects.

13. The computer-readable medium of claim 12 , wherein the instructions for predicting the forecast score for a test offer further cause the processor to:

determine a lift and standard deviation for the test offer based on the reinforcement learning model.

14. The computer-readable medium of claim 13 , wherein the forecast scores are a baseline function of time from the transaction log data plus elasticity from cross store experiments times a change in price, where in the elasticity is calculated as a function of the lift, and a confidence for the forecast score is calculated as a function of the standard deviation.

15. A system comprising a processor and a non-transitory computer-readable medium storing instructions that, when executed by the processor, cause the processor to perform the steps of:

receiving, at an offer generation system, data describing an offer on a product;

applying a natural language processing (NLP) model to the received data to extract the product and the offer;

accessing a plurality of test offers stored by an offer bank;

accessing transaction logs related to the product;

scoring each test offer in the offer bank against the extracted product, wherein scoring a test offer of the plurality of test offers comprises:

predicting a forecast score for each test offer by:

applying a reinforcement learning model to transaction logs related to the product, wherein the reinforcement learning model is trained using transaction logs associated with the plurality of test offers to predict a likelihood that a test offer will achieve an offer objective; and

determining a difference between the offer and the test offers; and

updating the forecast score by applying a penalty to the forecast score based on the determined difference;

selecting a subset of the plurality of test offers based in part on the forecast score of each test offer, wherein selecting the subset of test offers comprises maximizing an orthogonality of a set of variables associated with the product;

transmitting the subset of test offers to client devices of users;

receiving, from client devices of the users, responses to the subset of test offers, the responses comprising at least one of (1) whether or not a user viewed a test offer, (2) whether or not a user clicked through a test offer, (3) whether or not a user deleted a test offer, (4) whether or not the user saved the test offer to an electronic coupon folder, (5) whether or not a user forwarded the test offer to someone else, (6) whether or not a user posted a test offer to an online forum, or (7) whether or not the user purchased a product using the test offer;

storing the received responses in a database; and

retraining the reinforcement learning model based on the responses stored in the database, such that the reinforcement learning model automatically learns from user responses and continuously improves effectiveness of selection of test offers.

16. The system of claim 15 , wherein the instructions for assigning the subset of test offers to a plurality of retail locations causes the processor to:

deploy the offer and the selected subset of test offers in a plurality of retail locations.

17. The system of claim 16 , wherein the deploying is performed to maximize orthogonality between the following variables: store sales, store out of stock rates, number of relevant SKUs carried in each store, temporal effects, discount depth, buy quantity and offer structure.

18. The system of claim 15 , wherein the reinforcement learning model uses Thompson sampling.

19. The system of claim 15 , wherein the instructions for predicting a forecast score for a test offer by applying a reinforcement learning model to transaction logs further causes the processor to:

adjust the transaction logs for compliance by a retail location in a plurality of retail locations, estimated out of stock events, normalized across stores to account for different store attributes, and adjusted for temporal effects.

20. The system of claim 19 , wherein the instructions for predicting the forecast score for a test offer further cause the processor to:

determine a lift and standard deviation for the test offer based on the reinforcement learning model.

Assignments (2)

Continuity (11)

References Cited (300)