Systems and methods for cold-start recommendation using largescale graph models

Patent metadata

Field	Value
Publication number	US-12524445-B2
Application number	US-202418946069-A
Country	US
Kind code	B2
Filing date	Nov 13, 2024
Priority date	Jan 31, 2023
Publication date	Jan 13, 2026
Grant date	Jan 13, 2026

Abstract

Systems and methods of generating interfaces including recommended items selected by a graph-based cold-start (GCS) model are disclosed. A request for an interface is received and a set of interface items is generated for inclusion in the interface. The set of interface items is selected, at least in part, by a GCS model including a semantic similarity component and a viewed-also-viewed component. The set of interface items is generated based on a combination of an output of the semantic similarity component and an output of the viewed-also-viewed component. The interface including the set of interface items is generated and transmitted to a system that generated the request for the interface.

First claim

What is claimed is:

1 . A computer-implemented method, comprising:
receiving a request for an interface; generating, by a graph-based cold start (GCS) model, a sparse distance matrix and a co-viewed weighted matrix; generating a set of interface items for inclusion in the interface, wherein the set of interface items is selected based at least in part on the GCS model including a semantic similarity component and a viewed-also-viewed component, wherein the set of interface items is generated based on a combination of an output of the semantic similarity component and an output of the viewed-also-viewed component, wherein the semantic similarity component generates the sparse distance matrix and the viewed-also-viewed component generates the co-viewed weighted sparse matrix, and wherein the GCS model combines the sparse distance matrix and the co-viewed weighted sparse matrix to generate a product graph adjacency matrix; generating the interface including the set of interface items; and transmitting the interface to a system that generated the request for the interface.

2 . The computer-implemented method of claim 1 , wherein generating the sparse distance matrix comprises:
generating a distance matrix; and removing values above a first predetermined threshold or below a second predetermined threshold from the distance matrix.

3 . The computer-implemented method of claim 1 , wherein the sparse distance matrix includes cosine distance values for each item pair in a catalog of items.

4 . The computer-implemented method of claim 1 , comprising generating, by the GCS model, an adjacency graph based on the combination of the sparse distance matrix and the co-viewed weighted matrix, and wherein the set of interface items is selected at least in part by a graph traversal process that traverses the adjacency graph.

5 . The computer-implemented method of claim 4 , wherein the graph traversal process includes a Personalized PageRank process.

6 . The computer-implemented method of claim 1 , wherein the GCS model generates a set of candidate items, wherein generating the set of interface items comprises generating, by a ranking model, a set of ranked items by ranking the candidate items, and wherein the set of interface items is selected in descending rank order from the set of ranked items.

7 . The computer-implemented method of claim 1 , wherein the sparse distance matrix is representative of a set of cold-start item recommendations and the co-viewed weighted matrix is representative of a set of hot item recommendations.

8 . A non-transitory computer-readable medium storing instructions that, when executed by at least one processor, cause a device to execute operations comprising:
receiving a request for an interface; generating, by a graph-based cold start (GCS) model, a sparse distance matrix and a co-viewed weighted matrix; generating a set of interface items for inclusion in the interface, wherein the set of interface items is selected based at least in part on the GCS model including a semantic similarity component and a viewed-also-viewed component, wherein the set of interface items is generated based on a combination of an output of the semantic similarity component and an output of the viewed-also-viewed component, wherein the semantic similarity component generates the sparse distance matrix and the viewed-also-viewed component generates the co-viewed weighted sparse matrix, and wherein the GCS model combines the sparse distance matrix and the co-viewed weighted sparse matrix to generate a product graph adjacency matrix; generating the interface including the set of interface items; and transmitting the interface to a system that generated the request for the interface.

9 . The non-transitory computer-readable medium of claim 8 , wherein generating the sparse distance matrix comprises:
generating a distance matrix; and removing values above a first predetermined threshold or below a second predetermined threshold from the distance matrix.

10 . The non-transitory computer-readable medium of claim 8 , wherein the sparse distance matrix includes cosine distance values for each item pair in a catalog of items.

11 . The non-transitory computer-readable medium of claim 8 , comprising generating, by the GCS model, an adjacency graph based on the combination of the sparse distance matrix and the co-viewed weighted matrix, and wherein the set of interface items is selected at least in part by a graph traversal process that traverses the adjacency graph.

12 . The non-transitory computer-readable medium of claim 11 , wherein the graph traversal process includes a Personalized PageRank process.

13 . The non-transitory computer-readable medium of claim 8 , wherein the GCS model generates a set of candidate items, wherein generating the set of interface items comprises generating, by a ranking model, a set of ranked items by ranking the candidate items, and wherein the set of interface items is selected in descending rank order from the set of ranked items.

14 . The non-transitory computer-readable medium of claim 8 , wherein the sparse distance matrix is representative of a set of cold-start item recommendations and the co-viewed weighted matrix is representative of a set of hot item recommendations.

15 . A computer-implemented method, comprising:
receiving a request for an interface; generating, by a graph-based cold start (GCS) model, a sparse distance matrix and a co-viewed weighted matrix, wherein generating the sparse distance matrix includes:
generating a distance matrix; and
removing values above a first predetermined threshold or below a second predetermined threshold from the distance matrix, wherein the sparse distance matrix includes cosine distance values for each item pair in a catalog of items;
generating a set of interface items for inclusion in the interface, wherein the set of interface items is selected based at least in part on the GCS model including a semantic similarity component and a viewed-also-viewed component, wherein the set of interface items is generated based on a combination of an output of the semantic similarity component and an output of the viewed-also-viewed component, wherein the semantic similarity component generates the sparse distance matrix and the viewed-also-viewed component generates the co-viewed weighted sparse matrix, and wherein the GCS model combines the sparse distance matrix and the co-viewed weighted sparse matrix to generate a product graph adjacency matrix; generating the interface including the set of interface items; and transmitting the interface to a system that generated the request for the interface.

16 . The computer-implemented method of claim 15 , comprising generating, by the GCS model, an adjacency graph based on the combination of the sparse distance matrix and the co-viewed weighted matrix, and wherein the set of interface items is selected at least in part by a graph traversal process traverses the adjacency graph.

17 . The computer-implemented method of claim 16 , wherein the graph traversal process includes a Personalized PageRank process.

18 . The computer-implemented method of claim 15 , wherein the GCS model generates a set of candidate items, wherein generating the set of interface items comprises generating, by a ranking model, a set of ranked items by ranking the candidate items, and wherein the set of interface items is selected in descending rank order from the set of ranked items.

19 . The computer-implemented method

Assignees

• Walmart Apollo Llc

Inventors

• Bhardwaj Ankur
• Ali Afroza
• Achan Kannan

Classifications

• G06F16/9024

  Graphs; Linked lists (G06F16/9027 takes precedence) · CPC title

• G06F16/2237

  Vectors, bitmaps or matrices · CPC title

• G06F16/288 Primary

  Entity relationship models · CPC title