Method and apparatus for recommending content, device, and medium

Patent metadata

Field	Value
Publication number	US-12182729-B2
Application number	US-202117211720-A
Country	US
Kind code	B2
Filing date	Mar 24, 2021
Priority date	Sep 1, 2020
Publication date	Dec 31, 2024
Grant date	Dec 31, 2024

Abstract

A method and apparatus for recommending a content, a device, and a medium are provided. The method may include: determining, based on historical behavior data of a user using a product and a feature of a structure of a to-be-recommended content, a target structural preference of the user, the structure being determined by classifying the to-be-recommended content based on any classifying method of a content tag system; and determining each recommendation result of the user based on the target structural preference, the recommendation result including at least two structures and a recommendation content corresponding to each structure.

First claim

What is claimed is:

1. A method for recommending a content, comprising:
determining, in response to a session start request, a current session-level structural preference of a user using a pretrained session-level structural preference model based on historical behavior data of the user during a historical session and a feature of a structure of a to-be-recommended content, the structure being determined by classifying the to-be-recommended content based on any classifying method of a content tag system, a learning target of the session-level structural preference model comprising a preference degree of the user for any structure within a session, and the current session-level structural preference being used as a target structural preference of the user; and
determining each recommendation result of the user based on the target structural preference, the recommendation result comprising at least two structures and a recommendation content corresponding to each structure,
wherein a training process of the session-level structural preference model comprises:
using a plurality of training samples as model inputs, wherein each training sample is user behavior data generated during a session, each training sample comprises N sub-samples, each sub-sample is the user behavior data of each structure displayed for the session, and N is a natural number; and
using a percentage of a click rate of each structure involved in each pre-annotated training sample as a model output to train the session-level structural preference model.

2. The method according to claim 1 , wherein the historical behavior data comprises a historically selected structure content, a historically selected item content, a historically unselected structure content, and a historically unselected item content.

3. The method according to claim 1 , wherein the feature of the structure is represented by a structure name and a structure attribute, wherein the structure attribute represents historical click and display information of different structures.

4. The method according to claim 1 , wherein the determining the current session-level structural preference of the user based on the historical behavior data of the user during the historical session and the feature of the structure of the to-be-recommended content comprises:
determining the current session-level structural preference of the user based on the historical behavior data of the user during the historical session, a first scenario feature, and the feature of the structure of the to-be-recommended content, wherein the first scenario feature represents a scenario of each session.

5. The method according to claim 1 , further comprising:
determining, in response to each refresh request within the session, a current refresh-level structural preference of the user based on the current session-level structural preference and structured feedback information of the user, and using the current refresh-level structural preference as the target structural preference,
wherein the structured feedback information represents statisticized feedback information of different structures based on user behavior data within a historically set time period including a last refreshing process.

6. The method according to claim 4 , wherein the determining the current refresh-level structural preference of the user based on the current session-level structural preference and the structured feedback information of the user comprises:
adjusting the current session-level structural preference using an evolutionary learning model with the current session-level structural preference and the structured feedback information as model inputs, to output the current refresh-level structural preference.

7. The method according to claim 4 , wherein the determining the current refresh-level structural preference of the user based on the current session-level structural preference and the structured feedback information of the user comprises:
determining the current refresh-level structural preference of the user based on the current session-level structural preference, a second scenario feature, and the structured feedback information of the user, wherein the second scenario feature represents a scenario of each refresh.

8. The method according to claim 1 , further comprising, before the determining the each recommendation result of the user based on the target structural preference:
exploring in a structural dimension based on the target structural preference.

9. The method according to claim 1 , wherein the determining the each recommendation result of the user based on the target structural preference comprises:
determining the each recommendation result of the user through recalling, sorting, and fusion based on the target structural preference.

10. The method according to claim 9 , wherein the target structural preference comprises a weight representation of the user for each structural preference degree; and
accordingly, the determining the each recommendation result of the user through recalling, sorting, and fusion based on the target structural preference comprises:
directionally recalling, based on each structure in the target structural preference and a weight of the each structure, a content under the corresponding structure based on a recalling algorithm;
sorting the recalled content under the each structure based on a sorting algorithm; and
determining, based on the weight of the each structure, a position of the each structure in the recommendation result, and filling a corresponding content under the each structure into a corresponding structure based on a sorting result of the recalled content under the each structure, to obtain the recommendation result after the filling.

11. An electronic device, comprising:
at least one processor; and
a memory communicatively connected with the at least one processor;
the memory storing instructions executable by the at least one processor, and the instructions, when executed by the at least one processor, causing the at least one processor to perform operations, the operations comprising:
determining, in response to a session start request, a current session-level structural preference of a user using a pretrained session-level structural preference model based on historical behavior data of the user during a historical session and a feature of a structure of a to-be-recommended content, the structure being determined by classifying the to-be-recommended content based on any classifying method of a content tag system, a learning target of the session-level structural preference model comprising a preference degree of the user for any structure within a session, and the current session-level structural preference being used as a target structural preference of the user; and
determining each recommendation result of the user based on the target structural preference, the recommendation result comprising at least two structures and a recommendation content corresponding to each structure,
wherein a training process of the session-level structural preference model comprises:
using a plurality of training samples as model inputs, wherein each training sample is user behavior data generated during a session, each training sample comprises N sub-samples, each sub-sample is the user behavior data of each structure displayed for the session, and N is a natural number; and
using a percentage of a click rate of each structure involved in each pre-annotated training sample as a model output to train the session-level structural preference model.

12. A non-transitory computer readable storage medium storing computer instructions, the

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Liu Hesheng
• Wang Yu
• Xiong Yuling
• Peng Zhiming
• Song Naifei
• Qin Shouke

Classifications

• G06Q30/0631

  Recommending goods or services · CPC title

• G06N20/00

  Machine learning · CPC title

• H04N21/4532

  involving end-user characteristics, e.g. viewer profile, preferences (monitoring of user activities for profile generation for accessing a video database G06F16/739; user profiles in network data switching protocols H04L67/306; processing of user preferences or user profiles in wireless networks H04W8/18) · CPC title

• H04N21/4668

  for recommending content, e.g. movies · CPC title

• G06Q30/0255

  based on user history · CPC title