Granted Patent B2

US 12,586,117 · App. 18/357,648 · Granted Mar 24, 2026

Inventors: Bhavtosh Rath (Minneapolis, MN); Amit Pande (Ames, IA)

Assignee: Target Brands, Inc.

G06Q30/0631G06N3/042G06Q30/0633G06Q30/0603

Abstract

Claims (62)

1 . A method for providing session-based recommendations, the method comprising:

during an application session, receiving, at a user interface, selections of a first item and a second item by a user;

generating a first embedding for the first item;

generating a second embedding for the second item;

inferring, using a graph neural network trained to predict a next item in a sequence, a third item based on the selections of the first item and the second item, wherein inferring, using the trained graph neural network, the third item comprises:

generating a first graph comprising a first node and a second node;

assigning the first embedding to the first node;

assigning the second embedding to the second node;

modifying, using an edge-order preserving aggregation layer and a shortcut graph attention layer, the first embedding and the second embedding,

wherein using the edge-order preserving aggregation comprises preserving a selection order between the first item and the second item, and performing a first message passing to update the first embedding and the second embedding based on embeddings of neighboring nodes in the first graph;

wherein using the shortcut graph attention layer comprises adding, to the first graph, shortcut edges between pairs of nodes and subsequently selected nodes without an edge, and performing a second message passing to update the first embedding and the second embedding based on embeddings of neighboring nodes in the first graph;

after applying the edge-order preserving aggregation layer and the shortcut graph attention layer, generating a session embedding by combining the updated first embedding and the updated second embedding;

searching for a nearest embedding to the session embedding in an embeddings space;

during the application session, displaying, in the user interface, the third item, wherein the third item corresponds to the nearest embedding;

during the application session, receiving, at the user interface, a selection of the third item; and

during the application session, inferring, using the graph neural network, a fourth item based on the selection of the first item, the second item, and the third item wherein inferring, using the graph neural network comprises generating a second graph comprising a first node, a second node, and a third node corresponding to the first item, the second item, and the third item, respectively.

2 . The method of claim 1 ,

wherein searching for the nearest embedding comprises performing a comparison of the session embedding to a plurality of pre-computed item embeddings, the plurality of pre-computed item embeddings including a first set of pre-computed embeddings for items associated with the first item and a second set of pre-computed embeddings for items associated with the second item; and

wherein the method further comprises selecting, based at least in part on the comparison, one or more items from the plurality of items of the retail catalog, wherein the one or more items include the third item.

3 . The method of claim 2 , wherein generating the session embedding comprises applying an attention mechanism.

4 . The method of claim 1 , wherein the selections of the first item and the second item by the user are an ordered sequence.

5 . The method of claim 1 , wherein the third item is not a preselected recommendation for the selections of the first item and the second item.

6 . The method of claim 1 ,

wherein the graph neural network does not use biographical data of the user; and

wherein the graph neural network does not use historical activity data of the user.

7 . The method of claim 1 , wherein the application session is a browsing session or a mobile application session of the user.

8 . The method of claim 1 , wherein receiving the selections of the first item and the second item by the user includes receiving indications to add the first item and the second item to a shopping cart of the user.

9 . The method of claim 1 , wherein receiving the selections of the first item and the second item by the user includes receiving indications to view the first item and the second item.

10 . The method of claim 1 , wherein displaying the third item comprises presenting, in the user interface, the third item alongside a shopping cart of the user including the first item and the second item.

11 . A retail web server system configured to provide recommendations to a user interface accessible to one or more users, the retail web server system comprising:

at least one processor;

a memory subsystem including one or more memories, the memory subsystem storing instructions which, when executed by the at least one processor, cause the retail web server system to perform:

during an application session, receiving, at a user interface, selections of a first item and a second item by a user;

retrieving a first embedding generated for the first item;

retrieving a second embedding for the second item;

generating a first graph comprising a first node and a second node;

assigning the first embedding to the first node;

assigning the second embedding to the second node;

predicting, using a graph neural network trained to predict a next item in a sequence, a third item based on the selections of the first item and the second item, wherein predicting, using the trained graph neural network, the third item comprises modifying, using an edge-order preserving aggregation layer and a shortcut graph attention layer, the first embedding and the second embedding,

wherein using the edge-order preserving aggregation comprises preserving a selection order between the first item and the second item, and performing a first message passing to update the first embedding and the second embedding based on embeddings of neighboring nodes in the first graph;

wherein using the shortcut graph attention layer comprises adding, to the first graph, shortcut edges between pairs of nodes and subsequently selected nodes without an edge, and performing a second message passing to update the first embedding and the second embedding based on embeddings of neighboring nodes in the first graph;

after applying the edge-order preserving aggregation layer and the shortcut graph attention layer, generating a session embedding by combining the updated first embedding and the updated second embedding;

searching for a nearest embedding to the session embedding in an embeddings space;

during the application session, displaying, in the user interface, the third item, wherein the third item corresponds to the nearest embedding.

12 . The retail web server system of claim 11 , wherein the user interface is displayed on one or more of a mobile application or a website.

13 . The retail web server system of claim 11 , wherein displaying, in the user interface, the third item comprises replacing a previously recommended item with the third item.

14 . The retail web server system of claim 11 , wherein the instructions, when executed by the at least one processor, further cause the retail web server system to train the graph neural network using historical user activity data and synthetic simulation data.

15 . The retail web server system of claim 14 , wherein training the graph neural network comprises updating a weight used in the edge-order preserving aggregation layer and updating a weight used in the shortcut graph attention layer.

16 . The retail web server system of claim 11 , wherein the application session is a single browsing session of a single user.

17 . A method for providing session-based recommendations, the method comprising:

generating a first embedding for a first item using a graph neural network trained to predict a next item in a sequence;

generating a second embedding for a second item using the graph neural network;

during an application session, receiving, at a user interface, selections of the first item and the second item by a user;

generating a first graph comprising a first node and a second node;

assigning the first embedding to the first node;

assigning the second embedding to the second node;

predicting, using the graph neural network, a third item based on the selections of the first item and the second item, wherein predicting, using the trained graph neural network, the third item comprises modifying, using an edge-order preserving aggregation layer and a shortcut graph attention layer, the first embedding and the second embedding,

wherein using the edge-order preserving aggregation comprises assigning a selection order between the first item and the second item, and performing a first message passing to update the first embedding and the second embedding based on embeddings of neighboring nodes in the first graph;

wherein using the shortcut graph attention layer comprises adding, to the first graph, a shortcut edge between the first node and a subsequently selected node, and performing a second message passing to update the first embedding and the second embedding based on embeddings of neighboring nodes in the first graph;

after applying the edge-order preserving aggregation layer and the shortcut graph attention layer, generating a session embedding using the updated first embedding and the updated second embedding;

searching for a nearest embedding to the session embedding in an embeddings space; and

during the application session, displaying, in the user interface, the third item, wherein the third item corresponds to the nearest embedding.

Assignments (1)

Continuity (2)

References Cited (14)