US-20260127504-A1 - TIME-RESTRICTED MACHINE LEARNING MODELS

Abstract

A computer system receives a training data set that includes representations of a population of entities that are associated with a target action. Each representation in the training data set including a set of attributes of a respective entity and a time the target action associated with the respective entity occurred. The computer system segments the training data set into multiple time periods, then trains a model to identify, for each of the multiple time periods, an attribute that is correlated with a likelihood that the target action will occur with respect to an entity having the identified attribute during a corresponding time period. The computer system uses the model to perform a first operation for a first set of entities in a first time period and perform a second operation for a second set of entities in a second time period.

Inventors

• Jim SONG
• Sabitha VIKRAM
• Ketaki Bhalerao
• Andreas Hindman
• Srikanth Reddy ANNADI
• Winnie Chan

Assignees

• T-MOBILE USA, INC.

Dates

Publication Date

20260507

Application Date

20260105

Claims (20)

1. 1 . At least one computer-readable storage medium, excluding transitory signals and carrying instructions, which, when executed by at least one data processor of a system, cause the system to: select, using a model, a first set of users and a first time period based on a comparison between (1) a likelihood during a first time period that an action will occur with respect to a first set of users having a first attribute and (2) a likelihood during the first time period that the action will occur with respect to a second set of users that do not have the first attribute, wherein the model is trained, using a training data set, to cause the system to identify, for each of multiple time periods, one or more attributes that are correlated with a likelihood that the action will occur with respect to a user having the identified one or more attributes during a corresponding time period of the multiple time periods, and wherein the training data set comprises representations of users that are associated with the action, each representation in the training data set including a set of attributes of a respective user and a time, of a corresponding time period of the multiple time periods, at which the action associated with the respective user occurred; select, using the model, the second set of users and a second time period, of the multiple time periods, based on a comparison between (1) a likelihood during the second time period that the action will occur with respect to the second set of users and (1) a likelihood during the second time period that the action will occur with respect to the first set of users; perform a first operation for the selected first set of users during the first time period; and perform a second operation for the selected second set of users during the second time period.
2. 2 . The computer-readable storage medium of claim 1 , wherein the training data set is segmented into the multiple time periods by applying a rolling time window that selects a first subset of the representations that are associated with times falling within a first time period and selects a second subset of the representations that are associated with times falling within a second time period, and wherein the first and second time periods overlap.
3. 3 . The computer-readable storage medium of claim 1 , wherein the training data set is segmented into the multiple time periods by splitting the training data set into the first time period and a second time period that has a different length than the first time period.
4. 4 . The computer-readable storage medium of claim 1 , wherein the model is configured to output a ranking of multiple attributes of a population of users for each of the multiple time periods, and wherein the instructions when executed further cause the system to: select, as the first attribute, a highest-ranked attribute during the first time period.
5. 5 . The computer-readable storage medium of claim 1 , wherein the instructions further cause the system to select the first attribute and the first time period by: identifying one or more attributes for which the likelihood of the action occurring during the first time period for users having the identified one or more attributes is greater than an average likelihood of the action during the first time period; and selecting one of the identified one or more attributes as the first attribute.
6. 6 . The computer-readable storage medium of claim 1 , wherein a population of users, represented within the training data set, includes subscribers to a mobile network service, and wherein the action includes a subscriber canceling a subscription to the mobile network service, adding the subscription to the mobile network service, or modifying the subscription to the mobile network service.
7. 7 . The computer-readable storage medium of claim 6 , wherein the instructions further cause the processor to: select an incentive to send to subscribers having the first attribute during the first time period.
8. 8 . The computer-readable storage medium of claim 7 , wherein the first time period is a first portion of a year.
9. 9 . A method comprising: select, using a model, a first set of entities and a first time period based on a comparison between (1) a likelihood during a first time period that an action will occur with respect to a first set of entities having a first attribute and (2) a likelihood during the first time period that the action will occur with respect to a second set of entities that do not have the first attribute, wherein the model is trained, using a training data set, to cause a system to identify, for each of multiple time periods, one or more attributes that are correlated with a likelihood that the action will occur with respect to an entity having the identified one or more attributes during a corresponding time period of the multiple time periods, and wherein the training data set comprises representations of entities that are associated with the action, each representation in the training data set including a set of attributes of a respective entity and a time, of a corresponding time period of the multiple time periods, at which the action associated with the respective entity occurred; select, using the model, a second set of entities and a second time period, of the multiple time periods, based on a comparison between (1) a likelihood during the second time period that the action will occur with respect to the second set of entities and (2) a likelihood during the second time period that the action will occur with respect to the first set of entities; perform a first operation for the selected first set of entities during a first time period; and perform a second operation for the selected second set of entities during the second time period.
10. 10 . The method of claim 9 , wherein the training data set is segmented into the multiple time periods by applying a rolling time window that selects a first subset of the representations that are associated with times falling within a first time period and selects a second subset of the representations that are associated with times falling within a second time period, and wherein the first and second time periods overlap.
11. 11 . The method of claim 9 , wherein the training data set is segmented into the multiple time periods by splitting the training data set into the first time period and a second time period that has a different length than the first time period.
12. 12 . The method of claim 9 , wherein the model is configured to output a ranking of multiple attributes of a population of entities for each of the multiple time periods, and wherein the method further comprises: select, as the first attribute, a highest-ranked attribute during the first time period.
13. 13 . The method of claim 9 , further comprising selecting the first attribute and the first time period by: identifying one or more attributes for which the likelihood of the action occurring during the first time period for entities having the identified one or more attributes is greater than an average likelihood of the action during the first time period; and selecting one of the identified one or more attributes as the first attribute.
14. 14 . A system, comprising: at least one hardware processor; and at least one non-transitory memory storing instructions, which, when executed by the at least one hardware processor, cause the system to: select, using a model, a first set of entities and a first time period based on a comparison between (1) a likelihood during a first time period that an action will occur with respect to a first set of entities having a first attribute and (2) a likelihood during the first time period that the action will occur with respect to a second set of entities that do not have the first attribute, wherein the model is trained, using a training data set, to cause the system to identify, for each of multiple time periods, one or more attributes that are correlated with a likelihood that the action will occur with respect to an entity having the identified one or more attributes during a corresponding time period of the multiple time periods, and wherein the training data set comprises representations of entities that are associated with the action, each representation in the training data set including a set of attributes of a respective entity and a time, of a corresponding time period of the multiple time periods, at which the action associated with the respective entity occurred; select, using the model, a second set of entities and a second time period, of the multiple time periods, based on a comparison between (1) a likelihood during the second time period that the action will occur with respect to the second set of entities and (2) a likelihood during the second time period that the action will occur with respect to the first set of entities; perform a first operation for the selected first set of entities during the first time period; and perform a second operation for the selected second set of entities during the second time period.
15. 15 . The system of claim 14 , wherein the training data set is segmented into the multiple time periods by applying a rolling time window that selects a first subset of the representations that are associated with times falling within a first time period and selects a second subset of the representations that are associated with times falling within a second time period, and wherein the first and second time periods overlap.
16. 16 . The system of claim 14 , wherein the training data set is segmented into the multiple of time periods by splitting the training data set into the first time period and a second time period that has a different length than the first time period.
17. 17 . The system of claim 14 , wherein the model is configured to output a ranking of multiple attributes of a population of entities for each of the multiple time periods, and wherein the instructions when executed further cause the system to: select, as the first attribute, a highest-ranked attribute during the first time period.
18. 18 . The system of claim 14 , wherein the instructions further cause the system to select the first attribute and the first time period by: identifying one or more attributes for which the likelihood of the action occurring during the first time period for entities having the identified one or more attributes is greater than an average likelihood of the action during the first time period; and selecting one of the identified one or more attributes as the first attribute.
19. 19 . The system of claim 14 , wherein a population of entities, represented within the training data set, includes subscribers to a mobile network service, and wherein the action includes a subscriber canceling a subscription to the mobile network service, adding the subscription to the mobile network service, or modifying the subscription to the mobile network service.
20. 20 . The system of claim 19 , wherein the instructions further cause the system to: select an incentive to send to subscribers having the first attribute during the first time period.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application is a continuation of U.S. patent application Ser. No. 17/461,424, filed Aug. 30, 2021, which is hereby incorporated by reference in its entirety. BACKGROUND Machine learning is used in a wide variety of applications where development of conventional applications is difficult or unfeasible. Machine learning algorithms build a model based on sample data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so. For example, machine learning models are trained to identify features that, if present in an entity, likely result in a certain classification of the entity. However, conventional machine learning models do not take into account time-restricted variability in the training data, where for example some features are more relevant for classifying entities at some times and other features are more relevant at other times. As a result, these machine learning models are not able to make accurate predictions or decisions with respect to such time-varying data. BRIEF DESCRIPTION OF THE DRAWINGS Detailed descriptions of implementations of the present invention will be described and explained through the use of the accompanying drawings. FIG. 1 is a flowchart illustrating a process for generating time-restricted machine learning models. FIG. 2 is a graph illustrating example target action data. FIG. 3 is a block diagram that illustrates an example of a computer system in which at least some operations described herein can be implemented. The technologies described herein will become more apparent to those skilled in the art from studying the Detailed Description in conjunction with the drawings. Embodiments or implementations describing aspects of the invention are illustrated by way of example, and the same references can indicate similar elements. While the drawings depict various implementations for the purpose of illustration, those skilled in the art will recognize that alternative implementations can be employed without departing from the principles of the present technologies. Accordingly, while specific implementations are shown in the drawings, the technology is amenable to various modifications. DETAILED DESCRIPTION Implementations herein relate to generating and applying time-restricted machine learning models to discover the priority of event-dependent features in a population of entities. A computer system trains a time-restricted entity analysis model by segmenting a training data set into multiple time periods. The training data represents a population of entities that are associated with a target action, such as entities that performed the target action, and includes a set of attributes associated with each entity as well as a time the target action occurred with respect to each entity. For each of multiple time segments of the training data set, the computer system trains a model to identify an attribute that is correlated with a likelihood that the target action will occur with respect to an entity having the identified attribute during a corresponding time period. The computer system uses the trained model to select a first attribute and a first time period, which can be used, for example, to perform an operation related to the entities having the first attribute to reduce the likelihood of the target action before or during the first time period. Implementations of the time-restricted entity analysis model described herein enable the computer system to identify different attributes that are indicative, at different times, of whether the target action is likely to occur with respect to a population of entities. To illustrate time-restricted machine learning models, implementations are described herein with respect to an example application of identifying people who subscribe to mobile networks (“subscribers”) who are likely to cancel or modify their mobile network subscriptions. However, the time-restricted machine learning models described herein can be applied to any of a variety of types of time-varying with event specific characters within the data. The models described herein can be generated and used to analyze attributes of human and non-human entities in order to predict whether the entity is likely to perform any of a variety of target actions or have any of a variety of target actions occur with respect to the entity. For example, instead of predicting which people are likely to take a specified action at which times, similar models to those described herein can be used to predict the attributes of electronic systems that are most likely to fail at certain times or the attributes of people who are most likely to contract a disease at certain times. The description and associated drawings are illustrative examples and are not to be construed as limiting. This disclosure provides certain details for a thorough understanding and enabling description of these examples. One skille