Method for making recommendations to a user and apparatus, computing device, and storage medium

Patent metadata

Field	Value
Publication number	US-12250277-B2
Application number	US-202117329128-A
Country	US
Kind code	B2
Filing date	May 24, 2021
Priority date	Apr 18, 2019
Publication date	Mar 11, 2025
Grant date	Mar 11, 2025

Abstract

Embodiments of this application provide a method for making recommendations to a user and an apparatus, a computing device, and a storage medium. The method includes obtaining user attribute information, reading attribute information, reading history information, and candidate items; performing intra-group information fusion on the reading attribute information according to preset groupings to obtain reading feature information; obtaining a reading history weight according to the reading history information; obtaining history feature information according to the reading history weight and the reading history information; obtaining user feature information according to the user attribute information, the reading feature information, and the history feature information; and selecting a recommendation item from the candidate items according to the user feature information.

First claim

What is claimed is:

1. A method for making recommendations to a user, performed by a computing device, the method comprising:
obtaining user attribute information, reading attribute information, reading history information, and candidate items;
performing intra-group information fusion on the reading attribute information according to preset groupings to obtain reading feature information;
obtaining a reading history weight according to the reading history information;
obtaining history feature information according to the reading history weight and the reading history information;
obtaining user feature information according to the user attribute information, the reading feature information, and the history feature information;
inputting the user feature information and the candidate items into a neural network, to determine similarity scores that describe degree of similarity between the user feature information and the candidate items by using an inner product algorithm or a cosine similarity; and
selecting a recommendation item from the candidate items according to the similarity scores, wherein a quantity of the candidate items exceeds 10 million, and distributed k-nearest neighbor (k-NN) servers are provided to complete on-line real-time recall for selecting the recommendation item.

2. The method according to claim 1 , wherein the performing intra-group information fusion on the reading attribute information according to preset groupings to obtain reading feature information comprises:
performing average pooling on the reading attribute information in a same group to obtain the reading feature information.

3. The method according to claim 1 , wherein the obtaining a reading history weight according to the reading history information comprises:
inputting the reading history information to a self-attention layer model to calculate a self-attention layer output vector of the reading history information; and
calculating the reading history weight according to the self-attention layer output vector and a preset parameter.

4. The method according to claim 3 , wherein the obtaining history feature information according to the reading history weight and the reading history information comprises:
inputting the self-attention layer output vector to an attention pooling layer model, and performing weighted averaging on the self-attention layer output vector according to the reading history weight to obtain the history feature information.

5. The method according to claim 3 , wherein the self-attention layer model performs the following method:
re-encoding the reading history information to obtain a first feature matrix, a second feature matrix, and a third feature matrix;
performing linear transformation on the first feature matrix, the second feature matrix, and the third feature matrix, and inputting the first feature matrix, the second feature matrix, and the third feature matrix on which the linear transformation has been performed into a first self-attention model in parallel for a plurality of times, to obtain output results of the first self-attention model;
concatenating the output results of the first self-attention model and performing linear transformation to obtain a self-attention layer output vector.

6. The method according to claim 1 , wherein the obtaining user feature information according to the user attribute information, the reading feature information, and the history feature information comprises:
combining the user attribute information, the reading feature information, and the history feature information into combined feature information; and
inputting the combined feature information into a multi-layer neural network to obtain the user feature information that matches dimensions of the candidate items.

7. A user recommendation apparatus, comprising a processor and a memory, the memory storing computer-executable instructions, the computer-executable instructions, when executed by the processor, causing the processor to:
obtain corresponding reading feature information according to different reading attribute information;
obtain history feature information according to reading history information;
obtain user feature information according to user attribute information, the reading feature information, and the history feature information;
input the user feature information and the candidate items into a neural network, to determine similarity scores that describe degree of similarity between the user feature information and the candidate items by using an inner product algorithm or a cosine similarity; and
select a recommendation item from candidate items according to the similarity scores, wherein a quantity of the candidate items exceeds 10 million, and distributed k-nearest neighbor (k-NN) servers are provided to complete on-line real-time recall for selecting the recommendation item.

8. The apparatus according to claim 7 , wherein the computer-executable instructions further cause the processor to:
input the reading history information into a self-attention layer model to obtain a self-attention layer output vector;
calculate a reading history weight according to the self-attention layer output vector and a preset parameter; and
input the self-attention layer output vector to an attention pooling layer model, and perform weighted averaging on the self-attention layer output vector according to the reading history weight to obtain the history feature information.

9. The apparatus according to claim 8 , wherein obtaining the history feature information according to the reading history weight and the reading history information comprises:
inputting the self-attention layer output vector to an attention pooling layer model, and performing weighted averaging on the self-attention layer output vector according to the reading history weight to obtain the history feature information.

10. The apparatus according to claim 8 , wherein the self-attention layer model performs:
re-encoding the reading history information to obtain a first feature matrix, a second feature matrix, and a third feature matrix;
performing linear transformation on the first feature matrix, the second feature matrix, and the third feature matrix, and inputting the first feature matrix, the second feature matrix, and the third feature matrix on which the linear transformation has been performed into a first self-attention model in parallel for a plurality of times, to obtain output results of the first self-attention model;
concatenating the output results of the first self-attention model and performing linear transformation to obtain a self-attention layer output vector.

11. A non-transitory computer-readable storage medium, storing computer-executable instructions, the computer-executable instructions, when executed by a processor, causing the processor to perform the method for making recommendations to a user, comprising:
obtaining user attribute information, reading attribute information, reading history information, and candidate items;
performing intra-group information fusion on the reading attribute information according to preset groupings to obtain reading feature information;
obtaining a reading history weight according to the reading history information;
obtaining history feature information according to the reading history weight and the reading history information;
obtaining user feature information according to the user attribute information, the reading feature information, and the history feature information;
inputting the user feature information and the candidate items into a neural network, to determine similarity scores that describe degree of simil

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Qiu Zhijie
• Rao Jun
• Liu Yi
• Su Zhou
• Liu Shukai
• Sun Zhenlong
• Liu Qi
• Wang Liangdong
• Shang Tiantian
• Liang Mingfei
• Chen Lei
• Zhang Bo
• Lin Leyu

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/806

  of extracted features · CPC title

• G06N3/042

  Knowledge-based neural networks; Logical representations of neural networks · CPC title