Method for training classification model, classification method, apparatus and device

Patent metadata

Field	Value
Publication number	US-12282849-B2
Application number	US-202117349280-A
Country	US
Kind code	B2
Filing date	Jun 16, 2021
Priority date	Dec 28, 2020
Publication date	Apr 22, 2025
Grant date	Apr 22, 2025

Abstract

The present application discloses a method for training a classification model, a classification method, an apparatus and a device. A specific implementation is: acquiring behavior information of multiple users and personal basic information of the multiple users; where categories of at least part of users of the multiple users are known; inputting the personal basic information of the multiple users into a classification model to be trained to obtain feature information of the multiple users and predicted categories of users with known categories; and training the classification model to be trained according to the behavior information of the multiple users, the feature information of the multiple users, the predicted categories of the users with the known categories, and real categories of the users with the known categories, to obtain a trained classification model. The user categories determined by using the classification model are more accurate.

First claim

What is claimed is:

1. A method for training a classification model, comprising:
acquiring behavior information of multiple users and personal basic information of the multiple users; wherein categories of at least part of users of the multiple users are known;
inputting the personal basic information of the multiple users into a classification model to be trained to obtain feature information of the multiple users and predicted categories of users with known categories;
determining a first loss according to the behavior information of the multiple users and the feature information of the multiple users;
determining a second loss according to the predicted categories of the users with the known categories and real categories of the users with the known categories;
determining an end-to-end loss of the classification model to be trained according to the first loss and the second loss; and
training the classification model to be trained with a goal of minimizing the end-to-end loss to obtain a trained classification model; wherein in a training process, the feature information of the multiple users is updated by using the behavior information of the multiple users.

2. The method according to claim 1 , wherein the classification model to be trained comprises: a feature extracting network and a classification network;
the feature extracting network is configured to receive the personal basic information of the multiple users, perform feature extraction on the personal basic information of the multiple users, and output the feature information of the multiple users; and
the classification network is configured to receive feature information of the users with the known categories, classify the feature information of the users with the known categories, and output the predicted categories of the users with the known categories.

3. The method according to claim 2 , wherein the feature extracting network comprises: a similarity relation generating network and a feature generating network;
the similarity relation generating network is configured to receive the personal basic information of the multiple users, generate similarity relations among the multiple users according to the personal basic information of the multiple users, and output the similarity relations among the multiple users; and
the feature generating network is configured to receive the similarity relations among the multiple users, update the feature information of the multiple users according to the similarity relations among the multiple users, and output updated feature information of the multiple users.

4. The method according to claim 3 , wherein the similarity relation generating network comprises: multiple first neural networks and multiple second neural networks, wherein the first neural networks and the second neural networks are set at intervals;
the first neural networks are configured to receive the personal basic information of the multiple users, generate the similarity relations among the multiple users according to the personal basic information of the multiple users, and output the similarity relations among the multiple users to the second neural networks connected thereto, or to the feature generating network; and
the second neural networks are configured to receive the similarity relations among the multiple users, update the personal basic information of the multiple users according to the similarity relations among the multiple users, and output the updated personal basic information of the multiple users to the first neural networks connected thereto;
wherein the first neural networks are multilayer perceptron, and the second neural networks are graph convolutional networks.

5. The method according to claim 3 , wherein the feature generating network is a graph convolutional network.

6. The method according to claim 2 , wherein the classification network is a multilayer perceptron.

7. The method according to claim 2 , wherein the multiple users comprise users with unknown categories; after the trained classification model is obtained, the method further comprises:
inputting feature information of the users with the unknown categories into the classification network of the trained classification model to obtain predicted categories of the users with the unknown categories.

8. A classification method, comprising:
acquiring feature information of users with unknown categories;
inputting the feature information of the users with the unknown categories into a trained classification model, and determining predicted categories of the users with the unknown categories according to an output of the classification model;
wherein, the trained classification model is obtained by training according to behavior information of multiple users, personal basic information of the multiple users, and real categories of users with known categories; the multiple users comprise the users with the known categories and the users with the unknown categories, and the feature information of the users with the unknown categories is obtained by learning in a process of training the classification model, and the classification model is obtained by training by using the method according to claim 1 .

9. A classification apparatus, comprising:
at least one processor; and
a memory communicatively connected with the at least one processor; wherein,
the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, to enable the at least one processor to:
acquire feature information of users with unknown categories; and
input the feature information of the users with the unknown categories into a trained classification model, and determine predicted categories of the users with the unknown categories according to an output of the classification model;
wherein, the trained classification model is obtained by training according to behavior information of multiple users, personal basic information of the multiple users, and real categories of users with known categories; the multiple users comprise users with the known categories and the users with the unknown categories, where the feature information of the users with the unknown categories is obtained by learning in a process of training the classification model, and the classification model is obtained by training by using the method according to claim 1 .

10. An electronic device, including:
at least one processor; and
a memory communicatively connected with the at least one processor; wherein,
the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, to enable the at least one processor to execute the method according to claim 1 .

11. A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the method according to claim 1 .

12. A computer program product comprising a computer program, the computer program, when executed by a processor, implements the method according to claim 1 .

13. The method according to claim 1 , wherein in the training process, the obtained feature information of the multiple users that is infinitely close to the behavior information of the multiple users is taken as a training target, so as to improve accuracy of the feature information of the multiple users which is obtained by the classification model to be trained.

14. The method according to claim 1 , wherein the multiple users are multiple students, the behavior information of the multiple us

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Wang Yaqing
• Dou Dejing

Classifications

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06Q50/205

  Education administration or guidance · CPC title