Method and apparatus for training recommendation model, computer device, and storage medium

Patent metadata

Field	Value
Publication number	US-12544671-B2
Application number	US-202218076670-A
Country	US
Kind code	B2
Filing date	Dec 7, 2022
Priority date	Sep 15, 2020
Publication date	Feb 10, 2026
Grant date	Feb 10, 2026

Abstract

A recommendation model training method includes: acquiring first preference information corresponding to users for a first object, the first object being an object on which the users have performed a designated operation; determining second preference information corresponding to users for a second object, the second object being an object on which the users have not performed a designated operation; correcting the first and second preference information based on preference features of users for an object and correlation features between objects and the users, to obtain third preference information of the users for the objects; generating training data based on the third preference information, a user attribute of a target user, an object attribute of a target object, and operation information of the target user for the target object; and performing iterative training on a target recommendation model based on the training data until a training condition is satisfied.

First claim

What is claimed is:

1 . A method for training a recommendation model, performed by a computer device on a server, the method comprising:
acquiring first preference information corresponding to a plurality of users, the first preference information representing preference degrees of the users for a first object, the first object being an object on which the users have performed a designated operation; determining second preference information corresponding to the plurality of users based on similarities between the first preference information corresponding to the plurality of users, the second preference information representing preference degrees of the users for a second object, the second object being an object on which the users have not performed the designated operation, comprising:
respectively constructing preference vectors corresponding to the users based on the first preference information of the users for the first object;
respectively acquiring hash values of the preference vectors corresponding to the users;
grouping the users based on the preference vectors corresponding to the users, and categorizing at least two users into one group in response to that the hash values of the preference vectors corresponding to the at least two users are the same, wherein a user quantity of users included in each group is restricted so as to prevent an amount of data in a single group from exceeding a kernel memory of the server; and
determining, for one user, the second preference information of the user for the second object based on the first preference information corresponding to users belonging to the same group as the user;
correcting the first preference information and the second preference information based on preference features of users for an object and correlation features between objects and the users, to obtain third preference information of the users for the objects, the preference features and the correlation features being determined based on preference degrees of the users for the objects; generating, by a computing engine, training data based on the third preference information, a user attribute of a target user to which the third preference information belongs, an object attribute of a target object to which the third preference information belongs, and operation information of the target user for the target object, wherein before generating the training data, a storage type of a string format data is converted into a number format, so as to reduce memory consumption; and performing iterative training on a target recommendation model based on the training data, and obtaining a trained target recommendation model in response to that an output result of the target recommendation model satisfies a training condition.

2 . The method according to claim 1 , wherein the acquiring first preference information corresponding to a plurality of users comprises:
determining, for one user, a group of object use parameters corresponding to the user based on usage information of one first object by the user; acquiring weights corresponding to parameters in the group of object use parameters; and performing a weighting operation on the parameters in the group of object use parameters to obtain the first preference information of the user for the one first object.

3 . The method according to claim 2 , wherein the usage information of the one first object by the user comprises a first moment at which the user uses the one first object for the first time, a second moment at which the user uses the one first object for the last time, a total quantity of times that the user has used the one first object, and quantities of times that the user has used the one first object within at least two reference time periods.

4 . The method according to claim 3 , wherein the determining a group of object use parameters corresponding to the user based on usage information of one first object by the user comprises:
determining a time interval parameter based on the second moment and a current moment; determining a use frequency parameter based on the first moment, the current moment, and the total quantity of times; determining a preference parameter based on the first moment, the second moment, the current moment, and the total quantity of times, the preference parameter indicating a preference degree of the user for the one first object at the current moment; and determining a balancing factor based on the quantities of times that the user has used the one first object within the at least two reference time periods, the balancing factor being used for balancing preference degrees of the user for the one first object within the at least two reference time periods.

5 . The method according to claim 2 , wherein the acquiring weights corresponding to parameters in the group of object use parameters comprises:
respectively acquiring a priori scores corresponding to the parameters; and determining the weights corresponding to the parameters based on the a priori scores and usage information of first objects by the users.

6 . The method according to claim 1 , wherein after the determining similar users based on the similarities between the first preference information corresponding to the plurality of users, and categorizing the similar users into one group, the method further comprises:
in response to that a quantity of users comprised in one group is greater than a quantity threshold, selecting a first reference quantity of users from the group, and removing unselected users from the group.

7 . The method according to claim 1 , wherein the determining, for one user, the second preference information of the user for the second object based on the first preference information corresponding to users belonging to the same group as the user comprises:
for one user in a group of users, acquiring a second reference quantity of users with a similarity to the user satisfying a reference condition from the group of users as reference users; and determining the second preference information of the user for the second object based on the first preference information corresponding to the reference users.

8 . The method according to claim 1 , wherein the correcting the first preference information and the second preference information based on preference features of users for an object and correlation features between objects and the users, to obtain third preference information of the users for the objects comprises:
constructing a first matrix based on the first preference information and the second preference information corresponding to the users, one element in the first matrix representing preference information of one user for one object; performing M rounds of iterative update on the first matrix to obtain a second matrix of an M th round, M being a positive integer greater than or equal to 1, in a process of each round of iterative update, acquiring a second matrix obtained from a previous round of update as a new first matrix, performing matrix decomposition on the new first matrix to obtain the preference features of the users for the object and the correlation features between the objects and the users, and correcting elements in the new first matrix based on the preference features and the correlation features to obtain a new second matrix; and determining the third preference information based on elements in the second matrix of the M th round.

9 . The method according to claim 1 , wherein the generating training data based on the third preference information, a user attribute of a target user to which the third preference information belongs, an object attribute of a target object to which the third preference

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Wu Chongzheng
• Liu Biao
• Liu Dezhi
• Gan Quan
• Fei Qiang
• Shuai Pan
• Chen Ningguo
• Deng Jianwei
• Ke Xue

Classifications

• A63F2300/5506

  using advertisements · CPC title

• A63F13/61 Primary

  using advertising information · CPC title

• A63F13/79

  involving player-related data, e.g. identities, accounts, preferences or play histories · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title