US 12,423,722 B2
Incrementally updating embeddings for use in a machine learning model by accounting for effects of the updated embeddings on the machine learning model
Chuanwei Ruan, Sunnyvale, CA (US); Ramasubramanian Balasubramanian, Jersey City, NJ (US); and Peng Qi, Menlo Park, CA (US)
Assigned to Maplebear Inc., San Francisco, CA (US)
Filed by Maplebear Inc., San Francisco, CA (US)
Filed on Jul. 9, 2024, as Appl. No. 18/767,909.
Application 18/767,909 is a continuation of application No. 17/514,177, filed on Oct. 29, 2021, granted, now 12,051,081.
Prior Publication US 2024/0362657 A1, Oct. 31, 2024
This patent is subject to a terminal disclaimer.
Int. Cl. G06Q 30/00 (2023.01); G06Q 10/087 (2023.01); G06Q 30/0202 (2023.01)
CPC G06Q 30/0202 (2013.01) [G06Q 10/087 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
obtaining, at an online concierge system, an inventory of items offered by one or more warehouses, the inventory identifying a warehouse offering an item and attributes of the item;
generating, by the online concierge system, an item embedding for each item based on attributes of the item from the obtained inventory and to prior interactions by users of the online concierge system selecting items offered by at least one warehouse, the item embeddings being first latent space vectors in a first latent space;
generating, by the online concierge system, a user embedding for one or more users of the online concierge system based on characteristics of a user and to items the user previously purchased via the online concierge system, each user embedding being second latent space vectors in a second latent space;
obtaining new training data including a plurality of examples of items and of users that account for more recent interactions by users of the online concierge system with items;
generating updated embeddings the new training data, the updated embeddings including one or more item embeddings and/or one or more user embeddings;
retrieving a trained model configured to receive an input combination of a particular user and a particular item and to output a probability of the particular user performing a specific interaction with the particular item;
obtaining evaluation data including examples each comprising a combination of the item and the user with a label applied to each example indicating whether the user performed the specific interaction corresponding to the trained model to the item;
determining an existing error term for the trained model from application of the trained model to an item embedding of an item in an example and to a user embedding of a user in the example;
determining an updated error term for the trained model from application of the trained model to the one or more updated embeddings in the example;
responsive to determining a difference between the updated error term and the existing error term is not less than a threshold value:
adjusting at least one of the updated embeddings based on the updated error term and the existing error term;
retraining the trained model through adjusting one or more item embeddings in the first latent space and/or one or more user embeddings in the second space; and
storing one or more adjusted embeddings.