Granted Patent B2

US 12,547,645 · App. 18/514,236 · Granted Feb 10, 2026

Inventors: Wendi Liu (McLean, VA); Xue Li (Lewisville, TX); Fei Tong (Allen, TX); Bing Liu (Frisco, TX)

Assignee: Capital One Services, LLC

G06F16/285G06F16/116G06F16/182G06F16/1858G06F18/2193G06F18/2321G06F18/241G06F18/2413G06F21/554G06F21/561G06F21/71

Abstract

Claims (80)

1 . A system for classifying user files in a database into disparate states by mitigating bias using random data censoring, the system comprising:

one or more processors; and

a non-transitory, computer-readable medium comprising instructions that when executed by the one or more processors cause operations comprising:

receiving a request to determine respective probabilities that a user record for a user belongs to each of a plurality of user record states;

retrieving a user record history for the user record, wherein the user record history comprises periodic status checks, wherein each of the periodic status checks comprises a status check date;

determining that the user record history comprises a bias associated with the periodic status checks;

in response to determining that the user record history comprises a bias, generating a re-censored record history by randomly re-censoring the user record history based on a set build date, wherein generating the re-censored record history by re-censoring the user record history based on the set build date comprises:

determining the set build date based on an availability of data in the user record history;

determining a respective periodic status check proximity at each of the periodic status checks by:

retrieving a time stamp for each periodic status check of the periodic status checks; and

comparing the time stamp to the set build date;

determining a respective re-censoring probability at each of the periodic status checks based on the respective periodic status check proximity; and

determining whether a respective user record status at each of the periodic status checks is available based on the respective periodic status check proximity;

generating a feature input for the user record history based on the respective user record status at each of the periodic status checks that is available in the re-censored record history;

processing the feature input using a first model;

receiving a first output of the first model indicating the respective probabilities that the user record for the user belongs to each of the plurality of user record states; and

generating a recommendation based on the first output.

2 . A method of classifying user files in a database into disparate states by mitigating bias using random data censoring, the method comprising:

receiving a request to determine respective probabilities that a user record for a user belongs to each of a plurality of user record states;

retrieving a user record history for the user record, wherein the user record history comprises periodic status checks, wherein each of the periodic status checks comprises a status check date;

determining that the user record history comprises a bias associated with the periodic status checks;

in response to determining that the user record history comprises a bias, generating a re-censored record history by randomly re-censoring the user record history based on a set build date, wherein generating the re-censored record history by re-censoring the user record history based on the set build date comprises:

determining the set build date based on an availability of data in the user record history;

determining a respective periodic status check proximity at each of the periodic status checks by:

retrieving a time stamp for each periodic status check of the periodic status checks; and

comparing the time stamp to the set build date;

determining a respective re-censoring probability at each of the periodic status checks based on the respective periodic status check proximity; and

determining whether a respective user record status at each of the periodic status checks is available based on the respective periodic status check proximity;

generating a feature input for the user record history based on the respective user record status at each of the periodic status checks that is available in the re-censored record history;

processing the feature input using a first model;

receiving a first output of the first model indicating the respective probabilities that the user record for the user belongs to each of the plurality of user record states; and

generating a recommendation based on the first output.

3 . The method of claim 2 , wherein generating the re-censored record history by re-censoring the user record history based on the set build date further comprises:

receiving a user input of a data cutoff date; and

determining the set build date prior to the data cutoff date.

4 . The method of claim 2 , wherein generating the re-censored record history by re-censoring the user record history based on the set build date further comprises:

determining an availability of data in the user record history;

determining a data cutoff date based on the availability of data; and

determining the set build date prior to the data cutoff date.

5 . The method of claim 2 , wherein generating the re-censored record history by re-censoring the user record history based on the set build date further comprises:

determining a data cutoff date based on stationarity of data in the user record history;

determining the data cutoff date based on the stationarity of data; and

determining the set build date prior to the data cutoff date.

6 . The method of claim 2 , further comprising:

determining a number of the periodic status checks in the user record history; and

selecting the first model from a plurality of models based on the number of the periodic status checks in the user record history corresponding to a first range.

7 . The method of claim 2 , wherein generating the re-censored record history by re-censoring the user record history based on the set build date further comprises:

receiving a user input of a time-dependent re-censoring mechanism; and

applying the time-dependent re-censoring mechanism to the user record history.

8 . The method of claim 7 , wherein applying the time-dependent re-censoring mechanism to the user record history further comprises:

receiving a first datapoint in the user record history; and

setting an event status for the first datapoint based on the time-dependent re-censoring mechanism.

9 . The method of claim 2 , wherein processing the feature input using the first model further comprises:

applying a parametric survival analysis to the feature input; and

determining the first output based on the parametric survival analysis.

10 . The method of claim 2 , wherein processing the feature input using the first model further comprises:

applying a Kaplan-Meier curve analysis to the feature input; and

determining the first output based on the Kaplan-Meier curve analysis.

11 . The method of claim 2 , wherein processing the feature input using the first model further comprises:

applying a Cox proportional hazards regression to the feature input; and

determining the first output based on the Cox proportional hazards regression.

12 . The method of claim 2 , wherein the first model is a stochastic probability model comprising a Markov chain.

13 . The method of claim 2 , wherein the first model uses transition probability matrices to forecast a probability of an action related to the user record.

14 . The method of claim 2 , wherein the first output is a probability distribution row vector.

15 . The method of claim 2 , wherein the plurality of user record states comprises a user record state corresponding to a charge-off of the user record.

16 . One or more non-transitory, computer-readable mediums comprising instructions that, when executed by one or more processors, cause operations comprising:

receiving a request to determine respective probabilities that a user record for a user belongs to each of a plurality of user record states;

retrieving a user record history for the user record, wherein the user record history comprises periodic status checks, wherein each of the periodic status checks comprises a status check date;

determining that the user record history comprises a bias associated with the periodic status checks;

in response to determining that the user record history comprises a bias, generating a re-censored record history by re-censoring the user record history based on a set build date, wherein generating the re-censored record history by re-censoring the user record history based on the set build date comprises:

determining the set build date based on an availability of data in the user record history;

determining a respective periodic status check proximity at each of the periodic status checks by:

retrieving a time stamp for each periodic status check of the periodic status checks; and

comparing the time stamp to the set build date;

determining a respective re-censoring probability at each of the periodic status checks based on the respective periodic status check proximity; and

determining whether a respective user record status at each of the periodic status checks is available based on the respective periodic status check proximity;

generating a feature input for the user record history based on the respective user record status at each of the periodic status checks that is available in the re-censored record history;

processing the feature input using a first model;

receiving a first output of the first model indicating the respective probabilities that the user record for the user belongs to each of the plurality of user record states; and

generating a recommendation based on the first output.

Assignments (1)

Continuity (1)

References Cited (29)