Granted Patent B2

US 12,333,395 · App. 18/406,829 · Granted Jun 17, 2025

Inventors: Ernst Wilhelm Spannhake, II (Canal Winchester, OH); Thomas Francis Gianelle (Colleyville, TX); Milan Shah (Plano, TX)

Assignee: Citibank, N.A.

G06N20/00

Abstract

Claims (86)

1. A system for using cohort-based predictions in clustered time-series data in order to detect rate-of-change events, the system comprising:

one or more processors; and

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

generating historical time-series training data that indicates historic rates-of-change over given time periods;

receiving, at a user interface of a client portal, a request for a time-series prediction;

receiving a first user profile for a first user, wherein the first user profile comprises a current state characteristic, wherein the first user profile comprises a digital representation associated with an identity of the first user;

determining a subset of state characteristics for generating a first feature input based on the current state characteristic, wherein the first feature input comprises a vector array of values;

populating the first user profile with the subset of state characteristics;

in response to receiving the first user profile, determining the first feature input based on the subset of state characteristics;

retrieving a plurality of cohort clusters, wherein the plurality of cohort clusters is generated by a first artificial intelligence model that is trained to cluster a plurality of separate time-series data streams into the plurality of cohort clusters through unsupervised hierarchical clustering, and wherein data in each of the plurality of separate time-series data streams is cleansed by reformatting or standardizing data prior to clustering the data to into the plurality of cohort clusters;

inputting the first feature input into a second artificial intelligence model, wherein the second artificial intelligence model is trained to select a subset of the plurality of cohort clusters from the plurality of cohort clusters based on the first feature input, and wherein each cohort cluster of the plurality of cohort clusters corresponds to a respective cohort of users having current state characteristics, and wherein training the second artificial intelligence model comprises training a convolutional neural network using unsupervised learning on the historical time-series training data;

receiving an output from the second artificial intelligence model; and

generating, based on the output, a synthetic profile for the first user comprising the time-series prediction, wherein the synthetic profile comprises historic time-series data corresponding to the subset of the plurality of cohort clusters.

2. A method for using cohort-based predictions in clustered time-series data in order to detect rate-of-change events, the method comprising:

generating historical time-series training data that indicates historic rates-of-change over given time periods;

receiving a first user profile for a first user, wherein the first user profile comprises a current state characteristic, and wherein the first user profile comprises a digital representation associated with an identity of a first user;

in response to receiving the first user profile, determining a first feature input based on the first user profile, wherein the first feature input comprises a vector array of values;

retrieving a plurality of cohort clusters, wherein the plurality of cohort clusters is generated by a first artificial intelligence model that is trained to cluster a plurality of separate time-series data streams into the plurality of cohort clusters, wherein data in each of the plurality of separate time-series data streams is cleansed by reformatting or standardizing data prior to clustering the data to into the plurality of cohort clusters;

inputting the first feature input into a second artificial intelligence model, wherein the second artificial intelligence model is trained to select a subset of the plurality of cohort clusters from the plurality of cohort clusters based on the first feature input, wherein each cohort cluster of the plurality of cohort clusters corresponds to a respective cohort of users having current state characteristics, and wherein training the second artificial intelligence model comprises training a convolutional neural network using unsupervised learning on the historical time-series training data;

receiving an output from the second artificial intelligence model; and

generating, based on the output, a synthetic profile for the first user, wherein the synthetic profile comprises a time-series prediction.

3. The method of claim 2 , wherein determining the first feature input based on the first user profile comprises:

determining a subset of current state characteristics for generating a first feature input based on the current state characteristic;

populating the first user profile with the subset; and

in response to receiving the first user profile, determining the first feature input based on the subset.

4. The method of claim 3 , wherein populating the first user profile with the subset further comprises:

crawling the Internet for a remote server comprising the subset; and

retrieving the subset from the remote server.

5. The method of claim 2 , wherein determining the first feature input based on the first user profile further comprises:

receiving a user selection of the current state characteristic; and

determining to use the current state characteristic for the first feature input based on the first feature input.

6. The method of claim 2 , further comprising:

receiving a second user profile;

in response to receiving the second user profile, determining a second feature input for the second artificial intelligence model based on the second user profile;

inputting the second feature input into the second artificial intelligence model;

receiving a different output from the second artificial intelligence model; and

selecting, based on the different output, a different time-series prediction.

7. The method of claim 2 , wherein generating the time-series prediction further comprises:

determining a gradient for the time-series prediction;

determining a magnitude of the gradient; and

identifying a predicted event for the time-series prediction based on the magnitude.

8. The method of claim 7 , further comprising:

determining an effect of the predicted event based on the magnitude; and

generating a user recommendation based on the effect.

9. The method of claim 2 , wherein generating the time-series prediction further comprises:

determining a scale factor for the time-series prediction; and

performing, based on the scale factor, a linear transformation to the time-series prediction to identify a predicted event for the time-series prediction.

10. The method of claim 2 , wherein the first artificial intelligence model is trained to cluster the plurality of separate time-series data streams into the plurality of cohort clusters through unsupervised hierarchical clustering into hierarchies of correlation-distances between separate time-series data streams.

11. The method of claim 2 , wherein training the first artificial intelligence model comprises:

generating a matrix of pairwise correlations corresponding to the plurality of separate time-series data streams; and

clustering the plurality of separate time-series data streams based on pairwise distances.

12. The method of claim 2 , further comprising:

receiving a first labeled feature input, wherein the first labeled feature input is labeled with a known cohort cluster for the first labeled feature input; and

training the second artificial intelligence model to classify the first labeled feature input with the known cohort cluster.

13. The method of claim 2 , wherein generating the time-series prediction further comprises:

generating an isobaric representation of the time-series prediction; and

identifying a predicted event for the time-series prediction in the isobaric representation.

14. The method of claim 2 , wherein generating the time-series prediction further comprises:

identifying a predicted event for the time-series prediction; and

generating an option to modify the current state characteristic based on the predicted event.

15. A non-transitory, machine-readable medium for using cohort-based predictions in clustered time-series data in order to detect rate-of-change events, comprising instructions that, when executed by one or more processors, cause operations comprising:

generating historical time-series training data that indicates historic rates-of-change over given time periods;

receiving a first user profile, wherein the first user profile comprises a digital representation associated with an identity of a first user;

in response to receiving the first user profile, determining a first feature input based on the first user profile, wherein the first feature input comprises a vector array of values;

retrieving a plurality of cohort clusters, wherein the plurality of cohort clusters is generated by a first artificial intelligence model that is trained to cluster a plurality of separate time-series data streams into the plurality of cohort clusters, and wherein data in each of the plurality of separate time-series data streams is cleansed by reformatting or standardizing data prior to clustering the data to into the plurality of cohort clusters;

inputting the first feature input into a second artificial intelligence model, wherein the second artificial intelligence model is trained to select a subset of the plurality of cohort clusters from the plurality of cohort clusters based on the first feature input, wherein each cohort cluster of the plurality of cohort clusters corresponds to a respective cohort of users having current state characteristics, and wherein training the second artificial intelligence model comprises training a convolutional neural network using unsupervised learning on the historical time-series training data;

receiving an output from the second artificial intelligence model;

generating, based on the output, an isobaric representation of a time-series prediction, wherein the isobaric representation is based on historic time-series data corresponding to the subset of the plurality of cohort clusters; and

identifying a predicted event for the time-series prediction in the isobaric representation.

16. The non-transitory, machine-readable medium of claim 15 , wherein determining the first feature input based on the first user profile comprises:

determining a subset of current state characteristics for generating a first feature input;

populating the first user profile with the subset; and

in response to receiving the first user profile, determining the first feature input based on the subset.

17. The non-transitory, machine-readable medium of claim 16 , wherein populating the first user profile with the subset further comprises:

crawling the Internet for a remote server comprising the subset; and

retrieving the subset from the remote server.

18. The non-transitory, machine-readable medium of claim 15 , wherein determining the first feature input based on the first user profile further comprises:

receiving a user selection of a current state characteristic; and

determining to use the current state characteristic for the first feature input based on the first feature input.

19. The non-transitory, machine-readable medium of claim 15 , wherein generating the time-series prediction further comprises:

determining a gradient for the time-series prediction;

determining a magnitude of the gradient; and

identifying a predicted event for the time-series prediction based on the magnitude.

20. The non-transitory, machine-readable medium of claim 19 , wherein the instructions further cause operations comprising:

determining an effect of the predicted event based on the magnitude; and

generating a user recommendation based on the effect.

Assignments (1)

Continuity (3)

References Cited (47)