Granted Patent B2

US 12,450,227 · App. 18/482,774 · Granted Oct 21, 2025

Inventors: Cynthia Rogers (Palo Alto, CA); William Pentney (San Francisco, CA); Eric Pollmann (Los Altos, CA); Muhammad Bilal Mahmood (San Francisco, CA)

Assignee: Amplitude, Inc.

G06F16/2379G06F16/254

Abstract

Claims (80)

1. A computer-implemented method comprising:

storing a plurality of event records corresponding to a respective plurality of user interactions with a graphical user interface for a website or an application in a persistent storage, the storing comprising:

receiving the plurality of event records;

grouping the plurality of event records and compressing the plurality of event records for storage in a compressed format as a compressed set of event records, wherein the compressed set of event records is stored in association with a file path; and

writing the file path of the compressed set of event records to a publish-subscribe queue;

ingesting, from the publish-subscribe queue, a stream of data comprising the compressed set of event records, wherein the compressed set of event records is accessed using the file path as retrieved from a publish/subscribe message, wherein the ingesting decompresses the compressed set of event records back into the plurality of event records and is performed in parallel to the storing;

determining that a first record from the plurality of event records includes a first anonymous identifier and a first known identifier;

adding a mapping between the first anonymous identifier and the first known identifier to an identifier resolution database;

determining that a second record from the plurality of event records includes a second anonymous identifier and no known identifier;

using the identifier resolution database to identify a second known identifier that is mapped to the second anonymous identifier;

updating the second record to include the second known identifier; and

including the first record and the updated second record in a training matrix for training a machine learning model to compute causal inferences associated with the plurality of user interactions.

2. The computer-implemented method of claim 1 , further comprising:

parsing the plurality of event records based on identifying one or more feature groups into a plurality of parsed event records;

identifying an event aggregation axis by breaking out a particular feature group from the one or more feature groups;

generating a composite key based on the event aggregation axis and the particular feature group;

aggregating the plurality of event records based on the composite key and a selected aggregation type into aggregated event records; and

including the aggregated event records in the training matrix,

wherein the one or more feature groups comprise one or more of a categorical type of feature group, a count type of feature group, or a static type of feature group.

3. The computer-implemented method of claim 2 , wherein the categorical type of feature group is a feature group with unique values, the count type of feature group is a feature group with only one type of value, and the static type of feature group is a categorical type of feature group with greater than a threshold number of unique values.

4. The computer-implemented method of claim 2 , wherein the composite key comprises one or more of a user identifier, a feature group name identifier, an entity identifier, or an hour identifier.

5. The computer-implemented method of claim 2 , wherein updating the second record to include the second known identifier allows user interactions that occur prior to a user logging in to the website or the application to be attributed to specific users when training the machine learning model.

6. The computer-implemented method of claim 2 , wherein the selected aggregation type is one of a count type, a categorical type, or a static type.

7. The computer-implemented method of claim 2 , wherein aggregating the plurality of event records based on the composite key and the selected aggregation type comprises:

when the selected aggregation type is a category type, generating a record comprising a first value, a last value, and all unique values of a categorical feature group in the plurality of parsed event records.

8. The computer-implemented method of claim 2 , wherein aggregating the plurality of event records based on the composite key and the selected aggregation type comprises:

when the selected aggregation type is a static type, generating a record comprising a last value of a static feature group in the plurality of parsed event records.

9. The computer-implemented method of claim 2 , wherein aggregating the plurality of event records based on the composite key and the selected aggregation type comprises:

in response to the selected aggregation type being a count type, generating a record comprising a count of a number of instances of a count feature group in the plurality of parsed event records.

10. The computer-implemented method of claim 2 , further comprising:

detecting an aggregation trigger event; and

aggregating the plurality of event records in response to the detection of the aggregation trigger event.

11. The method of claim 1 , further comprising:

determining that a third record from the plurality of event records includes a third anonymous identifier and no known identifier;

determining that the third anonymous identifier is not included in the identifier resolution database;

waiting until the third anonymous identifier has been added to the identifier resolution database;

using the identifier resolution database to identify a third known identifier that is mapped to the third anonymous identifier;

updating the third record to include the third known identifier, and

adding the third record to the training matrix.

12. A system, comprising:

one or more processors; and

a non-transitory storage medium coupled to the one or more processors and storing instructions which, when executed by the one or more processors, cause the one or more processors to perform operations comprising:

storing a plurality of event records corresponding to a respective plurality of user interactions with a graphical user interface for a website or an application in a persistent storage, the storing comprising:

receiving the plurality of event records;

grouping the plurality of event records and compressing the plurality of event records for storage in a compressed format as a compressed set of event records, wherein the compressed set of event records is stored in association with a file path; and

writing the file path of the compressed set of event records to a publish-subscribe queue;

ingesting, from the publish-subscribe queue, a stream of data comprising the compressed set of event records, wherein the compressed set of event records is accessed using the file path as retrieved from a publish/subscribe message, wherein the ingesting decompresses the compressed set of event records back into the plurality of event records and is performed in parallel to the storing;

determining that a first record from the plurality of event records includes a first anonymous identifier and a first known identifier;

adding a mapping between the first anonymous identifier and the first known identifier to an identifier resolution database;

determining that a second record from the plurality of event records includes a second anonymous identifier and no known identifier;

using the identifier resolution database to identify a second known identifier that is mapped to the second anonymous identifier;

updating the second record to include the second known identifier; and

including the first record and the updated second record in a training matrix for training a machine learning model to compute causal inferences associated with the plurality of user interactions.

13. The system of claim 12 , wherein the instructions, when executed by the one or more processors, further cause performance of operations comprising:

parsing the plurality of event records based on identifying one or more feature groups into a plurality of parsed event records;

identifying an event aggregation axis by breaking out a particular feature group from the one or more feature groups;

generating a composite key based on the event aggregation axis and the particular feature group;

aggregating the plurality of event records based on the composite key and a selected aggregation type into aggregated event records; and

including the aggregated event records in the training matrix,

wherein the one or more feature groups comprise one or more of a categorical type of feature group, a count type of feature group, or a static type of feature group.

14. The system of claim 13 , wherein the categorical type of feature group is a feature group with unique values, the count type of feature group is a feature group with only one type of value, and the static type of feature group is a categorical type of feature group with greater than a threshold number of unique values.

15. The system of claim 13 , wherein the composite key comprises one or more of a user identifier, a feature group name identifier, an entity identifier, or an hour identifier.

16. The system of claim 13 , wherein updating the second record to include the second known identifier allows user interactions that occur prior to a user logging in to be attributed to specific users for use when training the machine learning model.

17. The system of claim 13 , wherein the selected aggregation type is one of a count type, a categorical type, or a static type.

18. The system of claim 13 , wherein the instructions, when executed by the one or more processors, cause performance of operations comprising:

for a categorical feature group, generating a record comprising a first value, a last value, and all unique values of the categorical feature group in the plurality of parsed event records.

19. The system of claim 13 , wherein the instructions, when executed by the one or more processors, cause performance of operations comprising:

for a static feature group, generating a record comprising a last value of the static feature group in the plurality of parsed event records.

20. The system of claim 13 , wherein the instructions, when executed by the one or more processors, cause performance of operations comprising:

for a count feature group, generating a record comprising a count of a number of instances of the count feature group in the plurality of parsed event records.

21. The system of claim 13 , wherein the instructions, when executed by the one or more processors, further cause performance of operations comprising:

detecting an aggregation trigger event; and

aggregating the plurality of event records in response to the detection of the aggregation trigger event.

22. The system of claim 12 , the operations further comprising:

determining that a third record from the plurality of event records includes a third anonymous identifier and no known identifier;

determining that the third anonymous identifier is not included in the identifier resolution database;

waiting until the third anonymous identifier has been added to the identifier resolution database;

using the identifier resolution database to identify a third known identifier that is mapped to the third anonymous identifier;

updating the third record to include the third known identifier; and

adding the third record to the training matrix.

Assignments (2)

Continuity (2)

References Cited (72)