Domain adaptation

Patent metadata

Field	Value
Publication number	US-12254062-B2
Application number	US-202017003104-A
Country	US
Kind code	B2
Filing date	Aug 26, 2020
Priority date	Aug 26, 2020
Publication date	Mar 18, 2025
Grant date	Mar 18, 2025

Abstract

Embodiments of the present disclosure relate to a method, system, and computer program product for domain adaptation. According to the method, a source model of a source domain is obtained, where the source model is trained to generate a label indicating a predicted category of data from the source domain. A training sample from a target domain is obtained, where the training sample comprises training data from the target domain and a true label indicating a true category of the training data from the target domain. A first label is generated for the training data by using the source model. The first label indicates a predicted category of the training data. A target model of the target domain is trained based on the training data, the true label and the first label.

First claim

What is claimed is:

1. A computer-implemented method comprising:
obtaining, by one or more processors, a source model of a source domain, wherein the source model is trained to predict at least one first category of first data from the source domain;
obtaining, by the one or more processors, a target training sample from a target domain, wherein the target training sample comprises target training data from the target domain and a target true label indicating a target true category of the target training data;
generating, by the one or more processors, a first label for the target training data by using the source model, wherein the first label indicates a predicted category of the target training data, wherein the first label is a first soft label generated by the source model;
training, by the one or more processors, a target model of the target domain based on the target training data, the target true label, the first label, and a plurality of weights associated with the target true label and the first label, wherein the target model is trained to predict at least one second category of second data from the target domain, wherein training the target model comprises:
generating, by the one or more processors, a second soft label and a hard label for the target training data by using the target model;
determining, by the one or more processors, a loss function based on a first difference between the hard label and the target true label and a second difference between the first soft label and the second soft label labels; and
training, by one or more processors, the target model by minimizing the loss function.

2. The method of claim 1 , wherein obtaining the source model comprises:
obtaining, by the one or more processors, a group of source training samples from the source domain, wherein the group of source training samples comprises source training data from the source domain and a source true label indicating a source true category of the source training data from the source domain; and
training, by the one or more processors, the source model based on the group of source training samples.

3. The method of claim 1 , wherein determining the loss function comprises:
weighting, by the one or more processors, the first difference based on a first weight and the second difference based on a second weight; and
determining, by the one or more processors, the loss function by summing the weighted first difference and the weighted second difference.

4. The method of claim 3 , further comprising:
in response to the first label matching the target true label, decreasing, by the one or more processors, the first weight while increasing the second weight.

5. The method of claim 3 , further comprising:
in response to the first label not matching the target true label, increasing, by the one or more processors, the first weight while decreasing the second weight.

6. The method of claim 1 , wherein the first data is first healthcare data of a first patient from a first hospital and the source model is trained to predict the at least one first category indicating a first predicted disease of the first patient, and where in the second data is second healthcare data of a second patient from a second hospital and the target model is trained to predict the at least one second category indicating a second predicted disease of the second patient.

7. A system comprising:
a processing unit; and
a memory coupled to the processing unit and storing instructions thereon, the instructions, when executed by the processing unit, performing actions comprising:
obtaining, by one or more processors, a source model of a source domain, wherein the source model is trained to predict at least one first category of first data from the source domain;
obtaining, by the one or more processors, a target training sample from a target domain, wherein the target training sample comprises target training data from the target domain and a target true label indicating a target true category of the target training data;
generating, by the one or more processors, a first label for the target training data by using the source model, wherein the first label indicates a predicted category of the target training data, wherein the first label is a first soft label generated by the source model;
training, by the one or more processors, a target model of the target domain based on the target training data, the target true label, the first label, and a plurality of weights associated with the target true label and the first label, wherein the target model is trained to predict at least one second category of second data from the target domain, wherein training the target model comprises:
generating, by the one or more processors, a second soft label and a hard label for the target training data by using the target model;
determining, by the one or more processors, a loss function based on a first difference between the hard label and the target true label and a second difference between the first soft label and the second soft label labels; and
training, by one or more processors, the target model by minimizing the loss function.

8. The system of claim 7 , wherein obtaining the source model comprises:
obtaining a group of source training samples from the source domain, wherein each of the group of source training samples comprises source training data from the source domain and a source true label indicating a source true category of the source training data from the source domain; and
training the source model based on the group of source training samples.

9. The system of claim 7 , wherein determining the loss function comprises:
weighting the first difference based on a first weight and the second difference based on a second weight; and
determining the loss function by summing the weighted first difference and the weighted second difference.

10. The system of claim 9 , wherein the actions further comprise:
in response to the first label matching the target true label, decreasing the first weight while increasing the second weight.

11. The system of claim 9 , wherein the actions further comprise:
in response to the first label not matching the target true label, increasing the first weight while decreasing the second weight.

12. The system of claim 7 , wherein the first data is first healthcare data of a first patient from a first hospital and the source model is trained to predict the at least one first category indicating a first predicted disease of the first patient, and where in the second data is second healthcare data of a second patient from a second hospital and the target model is trained to predict at least one second category indicating a second predicted disease of the second patient.

13. A computer program product being tangibly stored on a non-transient non-transitory machine-readable medium and comprising machine-executable instructions, the instructions, when executed on a device, causing the device to perform actions comprising:
obtaining, by one or more processors, a source model of a source domain, wherein the source model is trained to predict at least one first category of first data from the source domain;
obtaining, by the one or more processors, a target training sample from a target domain, wherein the target training sample comprises target training data from the target domain and a target true label indicating a target true category of the target training data;
generating, by the one or more processors, a first label for the target training data by using the source model, wherein the first label indicates a predicted category of the target trai

Assignees

• IBM

Inventors

• Yu yi qin
• Zhao Shiwan
• Mei Jing
• Li Shao Chun
• Min Xu

Classifications

• G06N3/096

  Transfer learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06F18/2431

  Multiple classes · CPC title

• G06F18/22

  Matching criteria, e.g. proximity measures · CPC title

• G06N3/08

  Learning methods · CPC title