Content recommendation method and apparatus, electronic device, and storage medium

Patent metadata

Field	Value
Publication number	US-11893071-B2
Application number	US-202217726048-A
Country	US
Kind code	B2
Filing date	Apr 21, 2022
Priority date	Apr 7, 2020
Publication date	Feb 6, 2024
Grant date	Feb 6, 2024

Abstract

This application provides a content recommendation method and apparatus, an electronic device, and a storage medium. The content recommendation method includes obtaining content feedback information of a target object and content feature information of content that is to be recommended in response to a content recommendation request of the target object, the content feedback information comprising explicit feedback information and implicit feedback information and object portrait information of the target object; performing feature interaction according to the explicit feedback information and the implicit feedback information in the content feedback information, and obtaining behavior preference information; performing feature extraction based on the behavior preference information, the content feedback information, and the content feature information, and obtaining a predicted click-through rate (CTR); and determining, according to the predicted CTR, recommended content from the pieces of content that is to be recommended, and transmitting the recommended content to a terminal device.

First claim

What is claimed is:

1. A content recommendation method, performed by an electronic device, the method comprising:
obtaining content feedback information of a target object and content feature information of content that is to be recommended in response to a content recommendation request of the target object, the content feedback information comprising at least explicit feedback information and implicit feedback information of the target object on recommended content and object portrait information of the target object;
respectively inputting the explicit feedback information and the implicit feedback information into a trained prediction model including a deep feedback interaction component, and obtaining behavior preference information of the target object outputted by the deep feedback interaction component;
performing feature extraction on the behavior preference information, the content feedback information, and the content feature information using the trained prediction model, and obtaining a predicted click-through-rate (CTR) that is outputted by the prediction model and at which the target object clicks on the content that is to be recommended; and
determining, according to the predicted CTR, recommended content from the pieces of content that is to be recommended, and transmitting the recommended content to a terminal device corresponding to the target object.

2. The method according to claim 1 , wherein
the trained prediction model is trained according to a training sample data set labeled with predicted CTRs, and a training sample in the training sample data set comprising content feedback information of a sample object and content feature information of sample content.

3. The method according to claim 2 , wherein the implicit feedback information comprises a click history sequence and an unclick history sequence corresponding to the target object, and the explicit feedback information comprises an explicit feedback history sequence corresponding to the target object; and
the respectively inputting the explicit feedback information and the implicit feedback information into the deep feedback interaction component, and obtaining the behavior preference information of the target object outputted by the deep feedback interaction component further comprises:
respectively inputting the click history sequence, the unclick history sequence, and the explicit feedback history sequence corresponding to the target object into the deep feedback interaction component, and performing encoding and decoding processing on the click history sequence, the unclick history sequence, and the explicit feedback history sequence based on a transformer based on a self-attention mechanism in the deep feedback interaction component to obtain a corresponding high-order click vector, a corresponding high-order unclick vector, and a corresponding explicit feedback high-order vector;
performing feature interaction on the high-order click vector and the unclick history sequence using an attention mechanism, to obtain a first interaction vector corresponding to the target object, and performing feature interaction on the explicit feedback high-order vector and the unclick history sequence to obtain a second interaction vector corresponding to the target object; and
concatenating the high-order click vector, the high-order unclick vector, the explicit feedback high-order vector, the first interaction vector, and the second interaction vector to obtain a behavior preference feature vector of the target object, and using the behavior preference feature vector as the behavior preference information of the target object.

4. The method according to claim 3 , wherein before the performing encoding and decoding processing on the click history sequence, the unclick history sequence, and the explicit feedback history sequence based on a transformer based on a self-attention mechanism in the deep feedback interaction component to obtain a corresponding high-order click vector, a corresponding high-order unclick vector, and a corresponding explicit feedback high-order vector, the method further comprises:
respectively fusing the click history sequence, the unclick history sequence, and the explicit feedback history sequence with time information.

5. The method according to claim 2 , wherein the object portrait information comprises a plurality of object portrait feature fields of the target object, and the content feature information comprises a plurality of content portrait feature fields of the content that is to be recommended; and
before the performing feature extraction on the behavior preference information, the content feedback information, and the content feature information using the trained prediction model, the method further comprises:
separately embedding sparse feature vectors respectively corresponding to the object portrait feature fields and the content portrait feature fields to obtain dense feature vectors respectively corresponding to the object portrait feature fields and the content portrait feature fields.

6. The method according to claim 5 , wherein the prediction model further comprises a wide component, a factorization machine (FM) component, and a deep neural networks (DNN) component, and the behavior preference information of the target object is a behavior preference feature vector; and
the performing feature extraction on the behavior preference information, the content feedback information, and the content feature information using the trained prediction model, and obtaining the predicted CTR that is outputted by the prediction model and at which the target object clicks the content that is to be recommended further comprises:
learning weight contributions of different feature fields in the object portrait information and the content feature information based on the wide component, and obtaining a feature weight vector;
performing feature extraction on the behavior preference feature vector and the dense feature vectors based on the FM component, and obtaining a low-order interaction feature vector corresponding to the target object;
performing feature extraction on the behavior preference feature vector and the dense feature vectors based on the DNN component, and obtaining a high-order interaction feature vector corresponding to the target object; and
concatenating the feature weight vector, the low-order interaction feature vector, and the high-order interaction feature vector to a fully-connected layer, and determining the predicted CTR at which the target object clicks the content that is to be recommended through weighted summation.

7. The method according to claim 6 , wherein the performing feature extraction on the behavior preference feature vector and the dense feature vectors based on the FM component, and obtaining a low-order interaction feature vector corresponding to the target object further comprises:
respectively inputting the dense feature vectors and the behavior preference feature vector into the FM component, and extracting an interaction result between any two feature vectors in the dense feature vectors and the behavior preference feature vector using Hadamard product; and
generating the low-order interaction feature vector based on the interaction result between different feature vectors.

8. The method according to claim 2 , wherein the trained prediction model is trained in the following manner:
selecting training samples from the training sample data set, each of the training samples being labeled with a predicted CTR at which a sample object clicks sample content;
for any training sample, inputting content feedback information of a sample object and content feature information of sample content comprised in the traini

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Ling Cheng
• Wang yalong
• Wang Rui
• Xie Ruobing
• Xia Feng
• Lin Leyu

Classifications

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06F16/9536 Primary

  Search customisation based on social or collaborative filtering · CPC title

• G06F16/9535 Primary

  Search customisation based on user profiles and personalisation · CPC title

• G06N3/045

  Combinations of networks · CPC title