Generating user-specific incentives based on previous activity using machine-learned large language models (LLMS)

1 . A method comprising:

accessing user behavior data and incentive data obtained prior to a current time, wherein the obtained incentive data describes previous incentives offered to users of an online system and the obtained user behavior data describes a degree of user interaction of the users with the online system in response to receiving the previous incentives;

training a behavior prediction model configured to receive user behavior data for a user and an incentive, and further configured to output an incentive prediction using the obtained user behavior data and incentive data, the behavior prediction model configured as a transformer architecture, and wherein training the behavior prediction model further comprises:

obtaining at least a training entry for known user behavior data for a particular user and known incentive data describing contents of incentives the particular user was presented with, and known interaction data describing how the user interacted with the incentives at a later time,

computing a loss function indicating a difference between estimated outputs and the known interaction data, wherein the estimated outputs are generated by applying the behavior prediction model to the known user behavior data and the known incentive data, and

backpropagating one or more terms obtained from the loss function to update parameters of the behavior prediction model;

receiving one or more candidate incentives generated by an incentive generation model for a particular user, wherein the incentive generation model is configured as another transformer architecture;

applying, to each candidate incentive and user behavior data for the particular user, the behavior prediction model to generate an incentive prediction describing a degree of interaction of the particular user with the online system in response to receiving the respective candidate incentive; and

transmitting, based on the incentive predictions determined for the one or more candidate incentives, instructions to cause display of a selected incentive to the particular user or another user on a client device.

2 . The method of claim 1 , wherein the selected incentive is a candidate incentive from the one or more candidate incentives that is associated with a highest degree of predicted interaction with the online system with respect to a desired objective for the particular user.

3 . The method of claim 1 , further comprising:

accessing a version of the incentive generation model with a fixed set of parameters;

generating one or more initial incentives using the version of the incentive generation model;

computing a loss function, wherein the loss function includes a first loss indicating a reward for the one or more candidate incentives obtained from the incentive predictions generated by the behavior prediction model, and a second loss indicating a divergence between content of the one or more initial incentives and the one or more candidate incentives; and

updating the parameters of the incentive generation model to reduce the loss function.

4 . The method of claim 1 , wherein at least one or both of the behavior prediction model and the incentive generation model is configured as a generative transformer architecture with one or more attention layers.

5 . The method of claim 1 , wherein the incentive prediction for each candidate incentive indicates one or a combination of an expected frequency of purchase by the particular user for a time period or a total amount of purchase by the particular user for a time period.

6 . The method of claim 1 , further comprising:

obtaining embeddings for the one or more candidate incentives mapping the one or more candidate incentives to a latent space; and

obtaining embeddings for the user behavior data for the particular user, and

wherein applying the behavior prediction model further comprises inputting the embeddings for the one or more candidate incentives and the embeddings for the user behavior data for the particular user to the behavior prediction model.

7 . The method of claim 1 , wherein the selected incentive is one or a combination of a coupon, an e-mail, or a marketing message.

8 . The method of claim 1 , further comprising:

obtaining interaction data of the particular user or the another user on the selected incentive;

generating additional training data from the interaction data; and

fine-tuning parameters of the behavior prediction model using the additional training data.

9 . A non-transitory computer-readable storage medium storing computer instructions, when executed by one or more processors, cause the one or more processors to:

access user behavior data and incentive data obtained prior to a current time, wherein the obtained incentive data describes previous incentives offered to users of an online system and the obtained user behavior data describes a degree of user interaction of the users with the online system in response to receiving the previous incentives;

train a behavior prediction model configured to receive user behavior data for a user and an incentive, and further configured to output an incentive prediction using the obtained user behavior data and incentive data, the behavior prediction model configured as a transformer architecture, wherein the instructions to train the behavior prediction model further cause the one or more processors to:

obtain at least a training entry for known user behavior data for a particular user and known incentive data describing contents of incentives the particular user was presented with, and known interaction data describing how the user interacted with the incentives at a later time,

compute a loss function indicating a difference between estimated outputs and the known interaction data, wherein the estimated outputs are generated by applying the behavior prediction model to the known user behavior data and the known incentive data, and

backpropagating one or more terms obtained from the loss function to update parameters of the behavior prediction model;

receive one or more candidate incentives generated by an incentive generation model for a particular user, wherein the incentive generation model is configured as another transformer architecture;

apply, to each candidate incentive and user behavior data for the particular user, the behavior prediction model to generate an incentive prediction describing a degree of interaction of the particular user with the online system in response to receiving the respective candidate incentive; and

transmit, based on the incentive predictions determined for the one or more candidate incentives, instructions to cause display of a selected incentive to the particular user or another user on a client device.

10 . The non-transitory computer-readable storage medium of claim 9 , wherein the selected incentive is a candidate incentive from the one or more candidate incentives that is associated with a highest degree of predicted interaction with the online system with respect to a desired objective for the particular user.

11 . The non-transitory computer-readable storage medium of claim 9 , wherein the computer instructions further cause the one or more processors to:

access a version of the incentive generation model with a fixed set of parameters;

generate one or more initial incentives using the version of the incentive generation model;

compute a loss function, wherein the loss function includes a first loss indicating a reward for the one or more candidate incentives obtained from the incentive predictions generated by the behavior prediction model, and a second loss indicating a divergence between content of the one or more initial incentives and the one or more candidate incentives; and

update the parameters of the incentive generation model to reduce the loss function.

12 . The non-transitory computer-readable storage medium of claim 9 , wherein at least one or both of the behavior prediction model and the incentive generation model is configured as a generative transformer architecture with one or more attention layers.

13 . The non-transitory computer-readable storage medium of claim 9 , wherein the incentive prediction for each candidate incentive indicates one or a combination of an expected frequency of purchase by the particular user for a time period or a total amount of purchase by the particular user for a time period.

14 . The non-transitory computer-readable storage medium of claim 9 , wherein the computer instructions further cause the one or more processors to:

obtain embeddings for the one or more candidate incentives mapping the one or more candidate incentives to a latent space; and

obtain embeddings for the user behavior data for the particular user, and

wherein applying the behavior prediction model further comprises inputting the embeddings for the one or more candidate incentives and the embeddings for the user behavior data for the particular user to the behavior prediction model.

15 . The non-transitory computer-readable storage medium of claim 9 , wherein the selected incentive is one or a combination of a coupon, an e-mail, or a marketing message.

16 . A computer system, comprising:

one or more processors; and

a non-transitory computer-readable storage medium storing computer instructions, when executed by one or more processors, cause the one or more processors to:

access user behavior data and incentive data obtained prior to a current time, wherein the obtained incentive data describes previous incentives offered to users of an online system and the obtained user behavior data describes a degree of user interaction of the users with the online system in response to receiving the previous incentives;

train a behavior prediction model configured to receive user behavior data for a user and an incentive, and further configured to output an incentive prediction using the obtained user behavior data and incentive data, the behavior prediction model configured as a transformer architecture, wherein the instructions to train the behavior prediction model further cause the one or more processors to:

obtain at least a training entry for known user behavior data for a particular user and known incentive data describing contents of incentives the particular user was presented with, and known interaction data describing how the user interacted with the incentives at a later time,

compute a loss function indicating a difference between estimated outputs and the known interaction data, wherein the estimated outputs are generated by applying the behavior prediction model to the known user behavior data and the known incentive data, and

backpropagating one or more terms obtained from the loss function to update parameters of the behavior prediction model;

receive one or more candidate incentives generated by an incentive generation model for a particular user, wherein the incentive generation model is configured as another transformer architecture;

apply, to each candidate incentive and user behavior data for the particular user, the behavior prediction model to generate an incentive prediction describing a degree of interaction of the particular user with the online system in response to receiving the respective candidate incentive; and

transmit, based on the incentive predictions determined for the one or more candidate incentives, instructions to cause display of a selected incentive to the particular user or another user on a client device.

17 . The computer system of claim 16 , wherein the selected incentive is a candidate incentive from the one or more candidate incentives that is associated with a highest degree of predicted interaction with the online system with respect to a desired objective for the particular user.

18 . The computer system of claim 16 , wherein the computer instructions further cause the one or more processors to:

access a version of the incentive generation model with a fixed set of parameters;

generate one or more initial incentives using the version of the incentive generation model;

compute a loss function, wherein the loss function includes a first loss indicating a reward for the one or more candidate incentives obtained from the incentive predictions generated by the behavior prediction model, and a second loss indicating a divergence between content of the one or more initial incentives and the one or more candidate incentives; and

update the parameters of the incentive generation model to reduce the loss function.

19 . The computer system of claim 16 , wherein at least one or both of the behavior prediction model and the incentive generation model is configured as a generative transformer architecture with one or more attention layers.

20 . The computer system of claim 16 , wherein the incentive prediction for each candidate incentive indicates one or a combination of an expected frequency of purchase by the particular user for a time period or a total amount of purchase by the particular user for a time period.