Machine learning module training using input reconstruction techniques and unlabeled transactions

Patent metadata

Field	Value
Publication number	US-11734558-B2
Application number	US-202016900129-A
Country	US
Kind code	B2
Filing date	Jun 12, 2020
Priority date	Jun 12, 2020
Publication date	Aug 22, 2023
Grant date	Aug 22, 2023

Abstract

Techniques are disclosed relating to improving machine learning classification using both labeled and unlabeled data, including electronic transactions. A computing system may train a machine learning module using a first set of transactions (of any classifiable data) with label information that indicates designated classifications for those transactions and a second set of transactions without label information. This can allow for improved classification error rates, particularly when additional labeled data may not be present (e.g., if a transaction was disallowed, it may not be later labeled as fraudulent or not). The training process may include generating first error data based on classification results for the first set of transactions, generating second error data based on reconstruction results for both the first and second sets of transactions, and updating the machine learning module based on the first and second error data.

First claim

What is claimed is:

1. A method for training machine learning models using input reconstruction results for unlabeled, incomplete transactions, comprising:
training, by a computing system, a machine learning module to classify electronic transactions, such that the machine learning module learns a characteristic of the electronic transactions that is indicative of different error data for those transactions, wherein the training uses a set of labeled transactions with labels indicating designated classifications for those transactions and a set of unlabeled transactions, and wherein the training includes:
generating first error data based on classification results generated by the machine learning module for the set of labeled transactions;
generating second, different error data based on reconstruction results generated via reconstruction of both the set of labeled transactions and the set of unlabeled transactions input into the machine learning module; and
updating the machine learning module based on the first and second error data; and

determining, by the computing system in response to receiving a request for authorization of a newly initiated transaction, to authorize the newly initiated transaction,
wherein the determining is performed using the machine learning module that is trained based on reconstruction results for both labeled and unlabeled transactions.

2. The method of claim 1 , wherein the set of unlabeled transactions are incomplete transactions and the set of labeled transactions are completed transactions.

3. The method of claim 1 , wherein transactions in the set of labeled transactions and the set of unlabeled transactions are electronic monetary transactions and wherein classifications output by the trained machine learning module specify whether transactions are authorized by an entity that had been requested to authorize those transactions.

4. The method of claim 1 , further comprising:
determining whether classification results for the set of unlabeled transactions match, within a threshold degree, with an expected distribution of classification results for unlabeled transactions.

5. The method of claim 4 , further comprising:
subsequent to the updating, further updating the machine learning module based on error data from the set of labeled transactions but not the set of unlabeled transactions based on determining that the classification results for the set of unlabeled transactions do not match the expected distribution.

6. The method of claim 1 , wherein the reconstruction results indicate whether a model implemented by the machine learning module is stale.

7. The method of claim 1 , further comprising:
tagging transactions in the set of unlabeled transactions to indicate that these transactions do not have label information, wherein one or more elements of the machine learning module are configured not to generate error data based on classification results for transactions of the set of unlabeled transactions based on detecting the tagging.

8. The method of claim 1 , further comprising:
determining whether classification results for the set of unlabeled transactions include a threshold percentage of results that have a particular classification; and
rolling back updates to the machine learning module, wherein the rolling back is performed based on determining that the classification results for the set of unlabeled transactions do not include the threshold percentage.

9. The method of claim 1 , wherein the reconstruction results are generated by generating output for the set of labeled transactions and the set of unlabeled transactions based on output of a first hidden layer of a neural network and not based on output of subsequent layers of the neural network.

10. A non-transitory computer-readable medium having instructions stored thereon that are executable by a computing device to perform operations comprising:
training a machine learning module to classify electronic transactions, such that the machine learning module learns a characteristic of the electronic transactions that is indicative of different error data for those transactions, wherein the training uses a set of labeled transactions with labels that indicate designated classifications for those transactions and a second, different set of unlabeled transactions, wherein the training includes:
generating first error data based on classification results generated by the machine learning module for the set of labeled transactions, wherein transactions in the set of labeled transactions are completed transactions;
generating second error data based on reconstruction results generated by an input reconstruction module for both the set of labeled transactions and the set of unlabeled transactions, wherein transactions in the set of unlabeled transactions are incomplete transactions; and
updating the machine learning module based on the first and second error data; and

determining, in response to receiving a request for authorization of a transaction, to authorize the transaction, wherein the determining is performed using the machine learning module that is trained based on reconstruction results for both labeled and unlabeled transactions.

11. The non-transitory computer-readable medium of claim 10 , wherein transactions in the sets of transactions are electronic monetary transactions, and wherein transactions in the set of labeled transactions are approved transactions and transactions in the set of unlabeled transactions are denied transactions.

12. The non-transitory computer-readable medium of claim 10 , wherein the operations further comprise:
determining whether classification results for the set of unlabeled transactions include a threshold percentage of results that have a particular classification.

13. The non-transitory computer-readable medium of claim 12 , wherein the operations further comprise:
rolling back updates to the machine learning module, wherein the rolling back is performed based on determining that the classification results for the set of unlabeled transactions do not include the threshold percentage.

14. The non-transitory computer-readable medium of claim 10 , wherein the operations further comprise:
tagging transactions in the set of unlabeled transactions to indicate that these transactions do not have label information, wherein one or more elements of the machine learning module are configured not to generate error data based on classification results for transactions of the set of unlabeled transactions based on detecting the tagging.

15. The non-transitory computer-readable medium of claim 10 , wherein the first error data indicates differences between the classification results and the designated classifications, and wherein the second error data indicates differences between the reconstruction results and input being reconstructed.

16. A method, comprising
receiving, by a computing system, a request for authorization of an electronic transaction;
accessing, by the computing system, a machine learning module that was trained such that the machine learning module learns a characteristic of the electronic transactions that is indicative of different error data for those transactions, wherein the machine learning model is trained prior to the accessing, by:
generating first error data based on classification results generated by the machine learning module for a first set of labeled transactions with labels indicating correct classifications for those transactions;
generating second, different error data based on reconstruction results generated via inpu

Assignees

• Paypal Inc

Inventors

• Saleh Moein
• Poletti Chiara
• Modaresi Sina
• Chen Yang
• Ji Xing

Classifications

• G06N3/0455

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

• 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

• G06N3/08 Primary

  Learning methods · CPC title