MACHINE LEARNING SYSTEM FOR TAKING CONTROL ACTIONS

• US 20200134628A1
• Filed: 10/26/2018
• Published: 04/30/2020
• Est. Priority Date: 10/26/2018
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Inventors: JIA, Yuting, OKA, Anand Ravindra, CHEN, Liang-Yu, WANG, Yiqing, NANDURI, Jayaram Naga Mrutyum
• Granted Patent: US 10,839,394 B2
• Time in Patent Office: Days
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/04 Inference or reasoning models
• CPC Class Codes: G06Q 20/4016 involving fraud or risk level assessment in transaction processing
• CPC Class Codes: G06Q 30/06 Buying, selling or leasing transactions

Abstract

A device in a data processing system for training machine learning models receives a transaction and forwards it to at least one of a plurality of integrated control action models that use outputs of one model as inputs to other models. The models are machine learning models jointly trained for taking each control action of a plurality of control actions on the transaction to maximize an objective function based on probabilities of the control actions matching corresponding target control actions. The machine learning models include a risk model that outputs risk prediction information for a first control action that indicates whether or not to initiate processing of the transaction. The device further receives the risk prediction information from the risk model, and executes at least the first control action based on the risk prediction information.

Claims

1. A method of training machine learning models, in a data processing system comprising at least one processor and at least one memory, the at least one memory comprising instructions executed by the at least one processor to process transactions, the method comprising:

receiving a transaction;

forwarding the transaction to at least one of a plurality of integrated control action models that use outputs of one model as inputs to other models, wherein the plurality of integrated control action models are machine learning models jointly trained for taking each control action of a plurality of control actions on the transaction to maximize an objective function based on a probability of the plurality of control actions matching corresponding target control actions taken on the transaction, wherein the plurality of integrated control action models include at least a risk model configured to output risk prediction information for a first control action that indicates whether or not to initiate processing of the transaction;

receiving the risk prediction information from the risk model; and

executing at least the first control action based on the risk prediction information.

2. The method of claim 1, wherein the plurality of integrated control action models further include a route model, wherein the route model outputs route prediction information for a second control action that indicates a processor to process the transaction if initiated, the method further comprising:

receiving the route prediction information from the route model; and

executing the second control action based on the risk prediction information and the route prediction information.

3. The method of claim 2, wherein the plurality of integrated merchant control action models further include a retry model, wherein the retry model outputs retry prediction information for a third control action that indicates whether or not to retry the transaction if declined, the method further comprising:

receiving declination information indicating that the transaction has been declined;

receiving the retry prediction information from the retry model; and

executing the third control action based on the risk prediction information, the route prediction information, and the retry prediction information.

4. The method of claim 3, wherein the plurality of integrated control action models further include a re-presentment model, wherein the re-presentment model outputs re-presentment prediction information for a fourth control action that indicates whether or not to re-present the transaction if a chargeback is initiated, the method further comprising:

receiving chargeback information indicating that the chargeback is initiated on the transaction;

receiving the re-presentment prediction information from the re-presentment model; and

executing the fourth control action based on the risk prediction information, the route prediction information, the retry prediction information, and the re-presentment prediction information.

5. The method of claim 1, wherein the objective function is an expected value of a profit associated with the transaction.

6. A method, in a data processing system comprising at least one processor and at least one memory, the at least one memory comprising instructions executed by the at least one processor to process transactions using machine learning, the method comprising:

receiving a current transaction; and

processing the current transaction by taking at least one control action using a joint learned model, wherein the joint learned model comprises at least a first learned model and a second learned model that are jointly trained for taking each control action of a plurality of control actions on at least one training transaction so as to maximize an expected value of an objective function associated with the at least one training transaction based on a probability of the plurality of control actions matching corresponding target control actions taken on the at least one training transaction using at least the first learned model and the second learned model.

7. The method of claim 6, wherein the second learned model is used to take a second control action on the at least one training transaction after the first learned model is used to take a first control action on the at least one training transaction, wherein an output of the first learned model is used as an input to the second learned model.

8. The method of claim 7, wherein data indicating whether a previously taken control action using the first learned model or using the second learned model has matched a corresponding target control action is used as an input to the first learned model or to the second learned model.

9. The method of claim 6, wherein the joint learned model comprises at least two learned models selected from the group consisting of a risk model configured to generate a risk score, a route model configured to generate a route score, a retry model configured to generate a retry score, and a re-presentment model configured to generate a re-presentment score.

10. The method of claim 9, wherein the risk score for the at least one training transaction is indicative of a probability of the at least one transaction being fraudulent, wherein the route score for the at least one training transaction is indicative of a probability of successful settlement of the at least one transaction using a routing schedule for routing the at least one training transaction or a payment processor to process the at least one training transaction, wherein the retry score for the at least one training transaction is indicative of a probability of successful settlement of the at least one transaction when the at least one training transaction has been declined and retried, wherein the re-presentment score for the at least one training transaction is indicative of a probability of success in re-presenting the at least one transaction when a chargeback has been issued on the at least one training transaction.

11. The method of claim 9, wherein the control actions taken on the at least one training transaction comprise at least two control actions selected from the group consisting of a risk control action taken based on the risk score, a route control action taken based on the route score, a retry control action taken based on the retry score, and a re-presentment control action taken based on the re-presentment score.

12. The method of claim 9, wherein the risk control action indicates whether to allow the at least one training transaction to proceed, wherein the route control action indicates a routing schedule for routing the at least one training transaction or a payment processor to process the at least one training transaction, wherein the retry control action indicates whether to retry the at least one training transaction if the at least one training transaction is declined, wherein the re-presentment control action indicates whether to comply with a chargeback request associated with the at least one training transaction.

13. The method of claim 9, wherein the risk score is provided as an input to at least one of the route model, the retry model, or the re-presentment model.

14. The method of claim 9, wherein the route score is provided as an input to at least one of the retry model or the re-presentment model.

15. The method of claim 9, wherein the retry score is provided as an input to the re-presentment model.

16. The method of claim 9, wherein at least one of the risk model, the route model, the retry model, or the re-presentment model is a real-time or near-real-time machine learning model.

17. The method of claim 9, wherein at least one of the risk model, the route model, the retry model, or the re-presentment model is a supervised or semi-supervised machine learning model.

18. The method of claim 9, wherein the re-presentment model further generates a re-presentment explanation configured to be provided as evidence for re-presenting the at least one training transaction.

19. The method of claim 18, wherein the re-presentment explanation is a human-understandable explanation.

20. A device in a data processing system for training machine learning models, comprising:

at least one processor; and

at least one memory in communication with the at least one processor, wherein the at least one memory comprises instructions executed by the at least one processor to process transactions, comprising;

receiving a transaction;

forwarding the transaction to at least one of a plurality of integrated control action models that use outputs of one model as inputs to other models, wherein the plurality of integrated control action models are machine learning models jointly trained for taking each control action of a plurality of control actions on the transaction to maximize an objective function based on a probability of the plurality of control actions matching corresponding target control actions taken on the transaction, wherein the plurality of integrated control action models include at least a risk model configured to output risk prediction information for a first control action that indicates whether or not to initiate processing of the transaction;

receiving the risk prediction information from the risk model; and

executing at least the first control action based on the risk prediction information.

21. A transaction processing apparatus, comprising:

a memory; and

at least one processor coupled to the memory, wherein the at least one processor is configured to process transactions using machine learning by;

receiving a current transaction; and

processing the current transaction by taking at least one control action using a joint learned model, wherein the joint learned model comprises at least a first learned model and a second learned model that are jointly trained for taking each control action of a plurality of control actions on at least one training transaction so as to maximize an expected value of an objective function associated with the at least one training transaction based on a probability of the plurality of control actions matching corresponding target control actions taken on the at least one training transaction using at least the first learned model and the second learned model.