US 12,093,979 B2
Systems and methods for generating real-time recommendations
Yokila Arora, Sunnyvale, CA (US); Gaoyang Wang, San Jose, CA (US); Shashank Kedia, Sunnyvale, CA (US); Shubham Gupta, Sunnyvale, CA (US); Aditya Mantha, San Jose, CA (US); Praveenkumar Kanumala, Newark, CA (US); Stephen Dean Guo, Saratoga, CA (US); and Kannan Achan, Saratoga, CA (US)
Assigned to Walmart Apollo, LLC, Bentonville, AR (US)
Filed by Walmart Apollo, LLC, Bentonville, AR (US)
Filed on Jan. 13, 2021, as Appl. No. 17/147,980.
Prior Publication US 2022/0222706 A1, Jul. 14, 2022
Int. Cl. G06Q 30/0251 (2023.01); G06Q 30/0601 (2023.01)
CPC G06Q 30/0256 (2013.01) [G06Q 30/0253 (2013.01); G06Q 30/0631 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A system comprising:
a non-transitory memory having instructions stored thereon; and
a processor configured to read the instructions to:
receive, from a user device, a search request including a user identifier associated with the search request, wherein the search request is generated during a current session;
receive user session data identifying one or more activities of the user from one or more servers, wherein the user session data is associated with the current session, and wherein the one or more activities are representative of interactions between the user device and the one or more servers;
receive historical user data associated with the user from a database;
generate a first set of embeddings based on the historical user data;
implement a trained favorite model to generate a first set of items from a plurality of items based on the first set of embeddings, wherein the trained favorite model is configured to apply a logit function to perform a pairwise comparison of each item in the plurality of items;
generate user intent for the current session based at least in part on the user session data, wherein the user intent is based at least in part on a set of items associated with the current user session;
generate a second set of embeddings based on the user session data;
train a context model based on pre-trained user embeddings and item embeddings, wherein the user embeddings are representative of a plurality of users during historical user sessions and the item embeddings are representative of items paired together during a corresponding historical user session;
implement the trained context model to generate a second set of items from the plurality of items based on the second set of embeddings and the user intent, the second set of items being different from the first set of items, and wherein the trained context model is configured to receive a data triplet including an embedding selected from the second set of embeddings and pre-trained feature embeddings associated with each of a first item and a second item;
implement a ranking model to generate a third set of items by re-ranking a combination of the first set of items and the second set of items, the third set of items including at least one item from the first set of items and at least one item from the second set of items, wherein the third set of items are determined by a trained ranking model configured to receive the first set of embeddings and the second set of embeddings; and
generate a user interface including the third set of items, wherein the user interface is presented via the user device.