Systems and methods for recommendation system analysis

Patent metadata

Field	Value
Publication number	US-12443983-B2
Application number	US-202217581635-A
Country	US
Kind code	B2
Filing date	Jan 21, 2022
Priority date	Jan 21, 2022
Publication date	Oct 14, 2025
Grant date	Oct 14, 2025

Abstract

Systems and methods including one or more processors and one or more non-transitory computer readable media storing computing instructions that, when executed on the one or more processors, cause the one or more processors to perform: receiving a user request via a graphical user interface, the user request corresponding to a user search query for a product; determining whether a first processing machine of the system is operating in a first processing mode or a second processing mode; when the first processing machine is determined to be operating in the first processing mode, analyzing the user request via the first processing machine and using a process, to identify a candidate recommendation system to utilize by: determining a randomized strategy for one or more candidate recommendation systems based on a ratio of a number of the one or more candidate recommender systems, the randomized strategy to be stored in a collected history data; determining model parameters based on the collected history data; and determining the candidate recommendation system from the one or more candidate recommendation systems as a candidate recommendation system with a maximum value for a reward model based on the user request; processing the user request with the candidate recommendation system to identify recommended products to display to the user; and transmitting instructions to modify the graphical user interface to display the recommended products to the user.

First claim

What is claimed is:

1. A system comprising:
one or more processors; and
one or more non-transitory computer-readable media storing computing instructions that, when executed on the one or more processors, cause the one or more processors to perform:
receiving a user request that is input to a web application included in an online component via a graphical user interface, the user request corresponding to a user search query from a user for a product;
determining, by a request distributor, included in the online component whether a system bandit service is operating in a first processing mode or a second processing mode, wherein the request distributor maintains computing resource availability and response time agreements, wherein the system bandit service operates as middleware between the online component and an offline component, wherein the system bandit service comprises a first distributed system with a first queue, a second distributed system with a second queue, an explore module, and an exploit module, wherein the request distributor determines which of the first queue or the second queue is a shorter queue with more memory, and wherein when the request distributor determines that the first queue is the shorter queue, the request distributor transmits the user request to the first queue of the first distributed system;
transmitting the user request to: (1) the exploit module, via the second distributed system, when the first processing mode is a high processing mode and the second processing mode is a low processing mode, or (2) the explore module, via the first distributed system, when the first processing mode is the low processing mode and the second processing mode is the high processing mode;
using a process of a reward model that comprises stage-wise exploration and exploitation with an optimal design, wherein the optimal design leverages a D-optimal design to increase information gain during exploration compared to not using the D-optimal design, by determining a randomized strategy for a plurality of candidate recommendation systems based on a ratio of a number of the plurality of candidate recommendation systems, the randomized strategy to be stored in a collected history data;
when the system bandit service is determined to be operating in the low processing mode, analyzing the user request, via the offline component, using the plurality of candidate recommendation systems and generating respective predicted reward values and corresponding pre-computations associated with the reward model, to be stored in the collected history data;

when the system bandit service is determined to be operating in the high processing mode, determining, via the system bandit service, the candidate recommendation system from the plurality of candidate recommendation systems as a recommendation system with a maximum reward value for the reward model based on the user request and the collected history data;
processing, via the system bandit service, the user request with one or more of the recommendation system or the one or more candidate recommendation systems to identify recommended products to display to the user; and
transmitting instructions, from the system bandit service to an algorithm service included in the online component, to modify the graphical user interface to display the recommended products to the user and determining a reward value for the one or more of the recommendation system or the one or more candidate recommendation systems, to be stored in the collected history data.

2. The system of claim 1 , wherein:
the first distributed system has the shorter queue than the second distributed system and the user request is transmitted to the first distributed system.

3. The system of claim 2 , wherein:
the first processing mode is the low processing mode;
the second processing mode is the high processing mode consuming more computing resources than the low processing mode; and
the process comprises an explore process.

4. The system of claim 2 , wherein:
the first processing mode is the high processing mode;
the second processing mode is the low processing mode consuming less computing resources than the high processing mode; and
the process comprises an exploit process.

5. The system of claim 1 , wherein:
the first processing mode is the low processing mode;
the second processing mode is the high processing mode consuming more computing resources than the low processing mode; and
the process comprises an explore process.

6. The system of claim 1 , wherein:
the first processing mode is the high processing mode;
the second processing mode is the low processing mode consuming less computing resources than the high processing mode; and
the process comprises an exploit process.

7. The system of claim 6 , wherein the exploit process further comprises resetting the collected history data collected prior to determining the randomized strategy.

8. The system of claim 1 , wherein determining the randomized strategy further comprises:
determining (a) a respective maximum value for a number of exploit rounds for each of the one or more candidate recommendation systems and (b) a respective maximum value for a number of explore rounds for each of the one or more candidate recommendation systems;
determining a respective potential randomized strategy based on a respective ratio for the number of exploit rounds and the number of explore rounds for each of the one or more candidate recommendation systems; and
determining a respective randomized strategy for each of the one or more candidate recommendation systems based on a relationship between the respective potential randomized strategy and a previous respective randomized strategy.

9. The system of claim 1 , further comprising:
determining model parameters included in the corresponding pre-computations associated with the reward model based on the collected history data comprising performing empirical-risk minimization.

10. A method implemented via execution of computing instructions configured to run at one or more processors and configured to be stored at non-transitory computer-readable media, the method comprising:
receiving a user request that is input to a web application included in an online component via a graphical user interface, the user request corresponding to a user search query from a user for a product;
determining, by a request distributor, included in the online component whether a system bandit service is operating in a first processing mode or a second processing mode, wherein the request distributor maintains computing resource availability and response time agreements, wherein the system bandit service operates as middleware between the online component and an offline component, wherein the system bandit service comprises a first distributed system with a first queue, a second distributed system with a second queue, an explore module, and an exploit module, wherein the request distributor determines which of the first queue or the second queue is a shorter queue with more memory, and wherein when the request distributor determines that the first queue is the shorter queue, the request distributor transmits the user request to the first queue of the first distributed system;
transmitting the user request to: (1) the exploit module, via the second distributed system, when the first processing mode is a high processing mode and the second processing mode is a low processing mode, or (2) the explore module, via the first distributed system, when the first processing mode is the low processing mode and the second processing mode is the high processing mode;
using a process of a reward model that comprises stage-wise e

Assignees

• Walmart Apollo Llc

Inventors

• Xu Da
• Xu Jianpeng
• Kumar Sushant
• Korpeoglu Evren
• Achan Kannan

Classifications

• G06Q30/0643

  graphically representing goods, e.g. 3D product representation · CPC title

• G06F9/505

  considering the load · CPC title

• G06Q30/0631 Primary

  Recommending goods or services · CPC title