Incremental time window procedure for selecting training samples for a supervised learning algorithm

Patent metadata

Field	Value
Publication number	US-11216751-B2
Application number	US-201916657450-A
Country	US
Kind code	B2
Filing date	Oct 18, 2019
Priority date	Oct 18, 2019
Publication date	Jan 4, 2022
Grant date	Jan 4, 2022

Abstract

Disclosed herein are system, method, and computer program product embodiments for generating labels for training a machine learning mode using an incremental time window process. The described process may be used in a recurrence detection system. A dataset may be analyzed using incremental split dates to divide the dataset into an analysis portion and a holdout portion. The analysis portion may be analyzed to determine input features related to a predicted recurrence in the dataset. The holdout portion may be tested against the analysis portion and the input features to generate a label. The label may indicate whether or not the holdout portion confirms the prediction. The testing of the holdout portion against the analysis portion may be repeated by incrementally using different split dates and multiple separate analysis portions and holdout portions to generate multiple labels and corresponding input features.

First claim

What is claimed is:

1. A computer-implemented method, comprising:
aggregating a dataset of data points;
performing a cadence analysis on the dataset to determine a recurrence period of the data points;
generating a first label using a first subset of the dataset having data points within a first multiple of the recurrence period;
generating a second label using a second subset of the dataset having a number of data points within a second multiple of the recurrence period, wherein the second subset includes the number of data points from the first subset within the first multiple of the recurrence period, and wherein the number of data points is based on a matching criteria that comprises a number of predictions and a date tolerance, and wherein the number of data points within the second multiple of the recurrence period are within the date tolerance and are equal to the number of predictions;
training a machine learning model using the first label and the second label;
receiving a set of transactions associated with an account and a merchant; and
generating, by the trained machine learning model, a predicted recurrence within the set of transactions, wherein generating the predicted recurrence further comprises:
determining a vector strength, a coverage, and a redundancy, wherein the vector strength, the coverage, and the redundancy are phase variables determined based on a mapping of transaction dates in a phase space associated with the set of transactions and characterizing a recurrence period within the set of transactions, and wherein the vector strength includes a first value reflecting a level of recurrence of the set of transactions within the recurrence period, the coverage includes a second value reflecting a number of recurrence periods that include no transactions from the set of transactions, and the redundancy includes a third value reflecting a number of recurrence periods that include at least one transaction from the set of transactions, wherein the set of transactions are plotted on the phase space for multiple recurrence periods with associated phase variables for each of the multiple recurrence periods and wherein the multiple recurrence periods includes the recurrence period; and
generating a probability of the merchant having the predicted recurrence within the set of transactions for the account, for the recurrence period, using the trained machine learning model based on an evaluation of at least one of the first value, the second value and the third value.

2. The computer-implemented method of claim 1 , wherein generating the first label further comprises:
designating the first subset as an analysis portion;
designating the data points having a time value exceeding the first multiple of the recurrence period as a holdout portion; and
testing the holdout portion against the analysis portion to generate the first label.

3. The computer-implemented method of claim 2 , further comprising: identifying a holdout multiple of the recurrence period; and
determining a delta between the holdout multiple and a data point of the dataset having a most recent time value to generate a holdout date.

4. The computer-implemented method of claim 3 , further comprising:
determining that a third multiple of the recurrence period exceeds the holdout date; and
in response to the determining, ceasing label generation.

5. The computer-implemented method of claim 1 , wherein the second multiple of the recurrence period is incrementally selected relative to the first multiple of the recurrence period.

6. The computer-implemented method of claim 1 , further comprising:
ceasing label generation in response to determining that no data points of the dataset have a time value between the second multiple of the recurrence period and a subsequent multiple of the recurrence period.

7. The computer-implemented method of claim 1 , further comprising:
generating a third label using a third subset of the data having data points within a third multiple of the recurrence period, wherein the third subset includes the data points from the first subset and the data points from the second subset.

8. A system, comprising:
a memory; and
at least one processor coupled to the memory and configured to:
aggregate a dataset of data points;
analyze the dataset to determine a recurrence period of the data points;
group the data points of the dataset into a first analysis subset and a first holdout subset, wherein the first analysis subset includes data points having a time value within the recurrence period;
generate a first label by testing the first holdout subset against the first analysis subset;
group the data points of the dataset into a second analysis subset and a second holdout subset, wherein the second analysis subset includes first data points from the first analysis subset and second data points having a time value within a multiple of the recurrence period, wherein the data points is based on a matching criteria that comprises a number of predictions and a date tolerance, and wherein the data points within the second analysis subset are within the date tolerance and are equal to the number of predictions;
generate a second label by testing the second holdout subset against the second analysis subset;
train a machine learning model using the first label and the second label;
receive a set of transactions associated with an account and a merchant; and
generate, by the trained machine learning model, a predicted recurrence within the set of transactions, wherein to generate the predicted recurrence further comprises:
determining a vector strength, a coverage, and a redundancy, wherein the vector strength, the coverage, and the redundancy are phase variables determined based on a mapping of transaction dates in a phase space associated with the set of transactions and characterizing a recurrence period within the set of transactions, and wherein the vector strength includes a first value reflecting a level of recurrence of the set of transactions within the recurrence period, the coverage includes a second value reflecting a number of recurrence periods that include no transactions from the set of transactions, and the redundancy includes a third value reflecting a number of recurrence periods that include at least one transaction from the set of transactions, wherein the set of transactions are plotted on the phase space for multiple recurrence periods with associated phase variables for each of the multiple recurrence periods and wherein the multiple recurrence periods includes the recurrence period; and
generating a probability of the merchant having the predicted recurrence within the set of transactions for the account, for the recurrence period, using the trained machine learning model based on an evaluation of at least one of the first value, the second value and the third value.

9. The system of claim 8 , wherein the difference in data points between the second analysis subset and the first analysis subset is the difference in data points between the first holdout subset and the second holdout subset.

10. The system of claim 8 , wherein the at least one processor is further configured to:
cease label generation in response to determining that no data points of the dataset have a time value between the multiple of the recurrence period and a subsequent multiple of the recurrence period.

11. The system of claim 8 , wherein the multiple of the recurrence period is incrementally selected relative to the recurrence period.

12. The system of claim 9 , wherein to generate the first label, the at least one processor is further configured to:
identify a se

Assignees

• Capital One Services Llc

Inventors

• Jumper Daniel
• Boroumand Jonathan
• GERSTLE Jeremy
• Zhao Jianshi

Classifications

• G06N20/00 Primary

  Machine learning · CPC title

• G06Q40/12

  Accounting · CPC title