Risk-based machine learning classsifier
The present disclosure relates generally to a risk-based fraud identification and risk analysis system. For example, the system may receive application data from a first borrower user, determine a segment associated with the application data, apply application data to one or more machine learning (ML) models, and receive a score based at least in part upon output of the ML model.
1. A method for computing an application score, the method comprising:
receiving, by a computer system, an application object for an application, wherein the application object includes application data associated with a first borrower user device, and wherein the application is initiated upon receiving a request from the first borrower user device at a second dealer user device or a third lender user device;
determining a segment associated with the application data, wherein determining a segment comprises:
extracting particular features from the application data, wherein the particular features comprise characteristics that group the application with other applications correlated with one of a borrower user, a dealer user, and a lender user; and
identifying a segment that corresponds to the particular features;
upon determining that the segment is a particular type of segment:
selecting a machine learning (ML) model associated with the segment from a plurality of trained ML models, wherein each of the trained ML models is dedicated to a respective segment;
generating, by the computer system, one or more input features associated with the application based upon the application data;
determining, by the computer system, a first output by applying the one or more input features associated with the application to the selected machine learning (ML) model;
determining, by the computer system, at least one second output by applying the one or more input features associated with the application to at least one ML model dedicated to a fraud type;
generating, by the computer system, a third output by combining the first output and the at least one second output;
scaling, by the computer system, the third output to a range of application scores to determine the application score for the application;
extracting adjustment features from the application data;
determining a discrepancy by comparing the adjustment features to threshold values associated with a standard user profile;
adjusting the application score based on the discrepancy;
determining, by the computer system, one or more reason codes for the application based at least in part on the application score for the application; and
determining, by the computer system, one or more actions for the application based at least in part on the adjusted application score for the application; and
providing, by the computer system, the adjusted application score, the one or more reason codes, and the one or more actions to the second dealer user device or the third lender user device.
2. The method for computing the application score of claim 1 , the method further comprising:
applying a weight to the one or more input features; and
training, by the computer system, the selected ML model using the weight applied to the one or more input features.
3. The method for computing the application score of claim 2 , the method further comprising:
adjusting the weight applied to the one or more input features based at least in part on a comparison to a current time, wherein more recent features have greater weight than less recent features.
4. The method for computing the application score of claim 2 , further comprising:
adjusting the weight based at least in part on discrepancies between a risk profile and the application data.
5. The method for computing the application score of claim 1 , wherein each of the trained ML models is trained using a training data set of historical application data.
6. A non-transitory computer-readable storage medium storing a plurality of instructions executable by one or more processors, the plurality of instructions when executed by the one or more processors cause the one or more processors to:
receive an application object for an application, wherein the application object includes application data associated with a first borrower user device, and wherein the application is initiated upon receiving a request from the first borrower user device at a second dealer user device or a third lender user device;
determine a segment associated with the application data, wherein determining a segment comprises:
extracting particular features from the application data, wherein the particular features comprise characteristics that group the application with other applications correlated with one of a borrower user, a dealer user, and a lender user; and
identifying a segment that corresponds to the particular features;
upon determining that the segment is a particular type of segment:
select a machine learning (ML) model associated with the segment from a plurality of trained ML models, wherein each of the trained ML models is dedicated to a respective segment;
generate one or more input features associated with the application based upon the application data;
determine a first output by applying the one or more input features associated with the application to the selected machine learning (ML) model;
determine at least one second output by applying the one or more input features associated with the application to at least one ML model dedicated to a fraud type;
generate a third output by combining the first output and the at least one second output;
scale the third output to a range of application scores to determine the application score for the application;
extract adjustment features from the application data;
determine a discrepancy by comparing the adjustment features to threshold values associated with a standard user profile;
adjust the application score based on the discrepancy;
determine one or more reason codes for the application based at least in part on the application score for the application; and
determine one or more actions for the application based at least in part on the adjusted application score for the application; and
provide the adjusted application score, the one or more reason codes, and the one or more actions to the second dealer user device or the third lender user device.
7. The non-transitory computer-readable storage medium of claim 6 , the one or more processors further caused to:
apply a weight to the one or more input features; and
train the selected ML model using the weight applied to the one or more input features.
8. The non-transitory computer-readable storage medium of claim 7 , the one or more processors further caused to:
adjusting the weight applied to the one or more input features based at least in part on a comparison to a current time, wherein more recent features have greater weight than less recent features.
9. The non-transitory computer-readable storage medium of claim 7 , the one or more processors caused to:
adjust the weight based at least in part on discrepancies between a risk profile and the application data.
10. The non-transitory computer-readable storage medium of claim 6 , wherein each of the trained ML models is trained using a training data set of historical application data.
11. A system comprising:
one or more processors; and
a non-transitory computer-readable medium including instructions that, when executed by the one or more processors, cause the one or more processors to:
receive an application object for an application, wherein the application object includes application data associated with a first borrower user device, and wherein the application is initiated upon receiving a request from the first borrower user device at a second dealer user device or a third lender user device;
determine a segment associated with the application data, wherein determining a segment comprises:
extracting particular features from the application data, wherein the particular features comprise characteristics that group the application with other applications correlated with one of a borrower user, a dealer user, and a lender user; and
identifying a segment that corresponds to the particular features;
upon determining that the segment is a particular type of segment:
select a machine learning (ML) model associated with the segment from a plurality of trained ML models, wherein each of the trained ML models is dedicated to a respective segment;
generate one or more input features associated with the application based upon the application data;
determine a first output by applying the one or more input features associated with the application to the selected machine learning (ML) model;
determine at least one second output by applying the one or more input features associated with the application to at least one ML model dedicated to a fraud type;
generate a third output by combining the first output and the at least one second output;
scale the third output to a range of application scores to determine the application score for the application;
extract adjustment features from the application data;
determine a discrepancy by comparing the adjustment features to threshold values associated with a standard user profile;
adjust the application score based on the discrepancy;
determine one or more reason codes for the application based at least in part on the application score for the application; and
determine one or more actions for the application based at least in part on the adjusted application score for the application; and
provide the adjusted application score, the one or more reason codes, and the one or more actions to the second dealer user device or the third lender user device.
12. The system of claim 11 , further comprising instructions that, when executed by the one or more processors, cause the one or more processors to:
apply a weight to the one or more input features; and
train the selected ML model using the weight applied to the one or more input features.
13. The system of claim 12 , further comprising instructions that, when executed by the one or more processors, cause the one or more processors to:
adjust the weight applied to the one or more input features based at least in part on a comparison to a current time, wherein more recent features have greater weight than less recent features.
14. The system of claim 12 , further comprising instructions that, when executed by the one or more processors, cause the one or more processors to:
adjust the weight based at least in part on discrepancies between a risk profile and the application data.
15. The non-transitory computer-readable storage medium of claim 11 , wherein each of the trained ML models is trained using a training data set of historical application data.