Generating recommendations using adversarial counterfactual learning and evaluation

Patent metadata

Field	Value
Publication number	US-12499375-B2
Application number	US-202117163398-A
Country	US
Kind code	B2
Filing date	Jan 30, 2021
Priority date	Jan 30, 2021
Publication date	Dec 16, 2025
Grant date	Dec 16, 2025

Abstract

A system 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, perform certain acts. The acts can include obtaining training data. The acts also can include training candidate recommendation models and an adversarial exposure model using the training data. The acts additionally can include generating recommendations based on a selected recommendation model of the candidate recommendation models. Other embodiments are described.

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:
generate first-stage click data by mimicking a generating mechanism of a real-world dataset;
obtain training data based on the first-stage click data and based on controlling an exposure mechanism in a second stage;
generate optimized machine learning models and a most adversarial exposure model by training a plurality of candidate recommendation models and an adversarial exposure model using the training data after obtaining the training data based on the first-stage click data and based on controlling the exposure mechanism in the second stage,
wherein the plurality of candidate recommendation models comprises multiple different families of machine learning models,
wherein training the adversarial exposure model comprises using sequential optimization, with two-timescale gradient ascent descent schema to optimize a minimax objective for the adversarial exposure model, by taking turns with model updating, and
wherein training the adversarial exposure model terminates when a validation metric changes less than a predetermined threshold after a predetermined number of consecutive epochs;

select, from the optimized machine learning models, a best optimized machine learning model by using the most adversarial exposure model to evaluate the optimized machine learning models on evaluation data; and
generate, based on receiving a request from a front-end component, recommendations by using the best optimized machine learning model.

2 . The system of claim 1 , wherein the training data comprise personalized recommendation data and item recommendation data.

3 . The system of claim 2 , wherein the personalized recommendation data comprise first records each comprising a respective user feature, a respective view sequence, a respective item feature, a respective target purchase, and a respective label.

4 . The system of claim 2 , wherein the item recommendation data comprise second records each comprising a respective anchor item, a respective candidate item, a respective item feature, and a respective label.

5 . The system of claim 1 , wherein the plurality of candidate recommendation models comprise a linear regression recommendation model, a neural network recommendation model, and a matrix factorization recommendation model.

6 . The system of claim 5 , wherein training the plurality of candidate recommendation models and the adversarial exposure model further comprises:
performing the two-timescale gradient ascent descent to optimize the minimax objective for each of the plurality of candidate recommendation models.

7 . The system of claim 1 , wherein, to select the best optimized machine learning model, the one or more processors are to:
perform an evaluation of the plurality of candidate recommendation models, as trained, using the adversarial exposure model, as trained.

8 . The system of claim 1 , wherein, to generate the recommendations, the one or more processors are to:
construct a recall set of candidate recommendation pairs; generate a ranking of candidate recommendation pairs in the recall set using the best optimized machine learning model; and determine the recommendations based on the ranking.

9 . The system of claim 1 , wherein the one or more processors are further to:
when a user requests to view an anchor item, sending one or more of the recommendations to be displayed to the user.

10 . A method implemented via execution of computing instructions configured to run at one or more processors, the method comprising:
generating first-stage click data by mimicking a generating mechanism of a real-world dataset; obtaining training data based on the first-stage click data and based on controlling an exposure mechanism; generating optimized machine learning models and a most adversarial exposure model by training a plurality of candidate recommendation models and an adversarial exposure model using the training data after obtaining the training data based on the first-stage click data and based on controlling the exposure mechanism,
wherein the plurality of candidate recommendation models comprises multiple different families of machine learning models,
wherein training the adversarial exposure model comprises using sequential optimization, with two-timescale gradient ascent descent schema to optimize a minimax objective for the adversarial exposure model, by taking turns with model updating, and
wherein training the adversarial exposure model terminates when a validation metric changes less than a predetermined threshold after a predetermined number of consecutive epochs;
selecting, from the optimized machine learning models, a best optimized machine learning model by using the most adversarial exposure model to evaluate the optimized machine learning models on evaluation data; and generating, based on receiving a request from a front-end component, recommendations by using the best optimized machine learning model.

11 . The method of claim 10 , wherein the training data comprise personalized recommendation data and item recommendation data.

12 . The method of claim 11 , wherein the personalized recommendation data comprise first records each comprising a respective user feature, a respective view sequence, a respective item feature, a respective target purchase, and a respective label.

13 . The method of claim 11 , wherein the item recommendation data comprise second records each comprising a respective anchor item, a respective candidate item, a respective item feature, and a respective label.

14 . The method of claim 10 , wherein the plurality of candidate recommendation models comprise a linear regression recommendation model, a neural network recommendation model, and a matrix factorization recommendation model.

15 . The method of claim 14 , wherein training the plurality of candidate recommendation models and the adversarial exposure model further comprises:
performing the two-timescale gradient ascent descent to optimize the minimax objective for each of the plurality of candidate recommendation models.

16 . The method of claim 10 , wherein selecting the best optimized machine learning model comprises:
performing an evaluation of the plurality of candidate recommendation models, as trained, using the adversarial exposure model, as trained; and selecting the best optimized machine learning model based on the evaluation.

17 . The method of claim 10 , wherein generating the recommendations comprises:
constructing a recall set of candidate recommendation pairs; generating a ranking of candidate recommendation pairs in the recall set using the best optimized machine learning model; and determining the recommendations based on the ranking.

18 . The method of claim 10 , further comprising:
when a user requests to view an anchor item, sending one or more of the recommendations to be displayed to the user.

19 . A non-transitory computer-readable medium storing computing instructions that, when executed on a processor, causes the processor to perform operations comprising:
generating first-stage click data by mimicking a generating mechanism of a real-world dataset; obtaining training data based on the first-stage click data and based on controlling an exposure mechanism; generating optimized machine learning mo

Assignees

• Walmart Apollo Llc

Inventors

• Xu Da
• Ruan Chuanwei
• Kumar Sushant
• Korpeoglu Evren
• Achan Kannan

Classifications

• G06N20/00

  Machine learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N3/045

  Combinations of networks · CPC title