Adaptive targeting for proactive voice notifications

1. A method comprising:

receiving, for a first item and a first user, first data comprising a first score representing a predicted likelihood that the first user will order the first item within a predefined amount of time;

determining a first set of features comprising features related to past voice notifications sent to the first user, past voice notifications consumed by the first user, and past orders of items associated with corresponding past voice notifications consumed by the first user;

determining a second set of features comprising features related to orders of the first item when a voice notification corresponding to an order of the first item was sent and orders of the first item when a voice notification corresponding to the order of the first item was not sent;

generating, by a first machine learning model, a first scaled score using the first score, the first set of features, and the second set of features;

determining, based on the first scaled score, that a first voice notification is to be sent to a device;

generating a first activity map comprising a first number of columns representing respective time slots and a first number of rows representing respective weeks, wherein a value of a first cell of the first activity map represents a first number of occurrences of a first event type during a first time slot and a first week associated with the first cell;

generating, by inputting the first activity map into a convolutional neural network comprising at least a first convolutional layer and a first pooling layer, first feature data comprising a down-sampled representation of the first activity map;

generating, by at least one fully-connected classification layer using the first feature data, a first vector, wherein each element of the first vector corresponds to a respective time slot, and wherein a respective value of each element of the first vector represents a respective predicted probability that an event of the first event type will occur if the first voice notification is sent to the device during the respective time slot;

selecting a second time slot of the first vector based on a first predicted probability associated with the second time slot in the first vector; and

sending audio data comprising the first voice notification to the device associated with the first user during the second time slot, wherein the first voice notification includes a prompt to remind the first user to reorder the first item.

2. The method of claim 1 , further comprising:

determining a second activity map comprising a third set of features representing a timing of past utterances of the first user directed to a natural language processing system;

determining a third activity map comprising a fourth set of features representing a timing at which a first subset of the past voice notifications sent to the first user were sent, wherein the first subset of the past voice notifications were consumed;

determining a fourth activity map comprising a fifth set of features representing a timing at which a second subset of the past voice notifications sent to the first user were sent, wherein the second subset of the past voice notifications were not consumed;

determining, by the convolutional neural network, a third time slot comprising a day of a week and an hour of the day, wherein the convolutional neural network determines the third time slot using at least the third set of features, the fourth set of features, and the fifth set of features; and

sending second audio data comprising a second voice notification to the device during the third time slot.

3. The method of claim 1 , further comprising:

determining, by a sensor associated with the device associated with the first user, second data indicating a presence of the first user at a location proximate to the device; and

selecting the device for outputting the audio data from among other devices associated with the first user based on the second data indicating the presence of the first user.

4. A method comprising:

receiving, for first content and a first user, first data representing a predicted likelihood that the first user will interact with the first content within a predefined amount of time;

determining a first set of features comprising features related to past voice notifications sent to the first user;

determining a second set of features comprising features related to interaction with the first content when one of the past voice notifications corresponding to the first content was sent;

generating, by a first machine learning model based at least in part on the first data, the first set of features, and the second set of features, a prediction that a voice notification will increase a probability that the first user will interact with the first content;

generating a first activity map comprising a first number of columns and a first number of rows, wherein a first cell at a first intersection of a first column and a first row corresponds to a first time slot during a first week, wherein a value of the first cell represents a first number of occurrences of a first event type during the first time slot and the first week;

generating, by inputting the first activity map into a convolutional neural network comprising at least a first convolutional layer and a first pooling layer, first feature data comprising a down-sampled representation of the first activity map;

generating, by at least one classification layer using the first feature data, a first vector, wherein a first element of the first vector corresponds to a second time slot, and wherein a value of the first element represents a predicted probability that an event of the first event type will occur if the voice notification is sent to a first device during the second time slot;

selecting the second time slot based on the predicted probability associated with the second time slot in the first vector; and

sending audio data comprising the voice notification to the first device during the second time slot, wherein the voice notification includes a prompt related to the first content.

5. The method of claim 4 , further comprising:

determining a first feature representing a subset of the past voice notifications that were consumed by the first user; and

determining a second feature representing a subset of the past voice notifications that were not consumed by the first user, wherein the first set of features comprises the first feature and the second feature.

6. The method of claim 4 , further comprising determining a first feature representing a subset of the past voice notifications that were consumed and which were associated with past interaction with content.

7. The method of claim 4 , further comprising:

determining a first feature representing orders associated with past voice notifications that were consumed by the first user; and

determining a second feature representing orders associated with past voice notifications that were not consumed by the first user, wherein the second set of features comprise the first feature and the second feature.

8. The method of claim 4 , further comprising determining, by the convolutional neural network, a time at which to send the audio data based at least in part on a timing of a subset of the past voice notifications that were consumed by the first user.

9. The method of claim 4 , further comprising determining, by the convolutional neural network, a time at which to send the audio data based at least in part on historical data indicating one or more times-of-day associated with the first user interacting with a natural language processing system.

10. The method of claim 4 , further comprising:

determining a third set of features representing a first frequency at which the past voice notifications were sent; and

determining, by a second machine learning model, a second frequency with which to send voice notifications based at least in part on the first set of features, the second set of features, and the third set of features.

11. The method of claim 4 , further comprising determining, by a second machine learning model, the first device from among other devices based at least in part on presence data indicating detection of the first user by a sensor associated with the first device.

12. The method of claim 11 , further comprising determining, by a third machine learning model a first time of day at which to send the audio data to the first device based at least in part on historical data indicating times at which the first user has historically consumed voice notifications.

13. A system comprising:

at least one processor; and

non-transitory computer-readable memory storing instructions that, when executed by the at least one processor, are effective to:

receive, for first content and a first user, first data representing a predicted likelihood that the first user will interact with the first content within a predefined amount of time;

determine a first set of features comprising features related to past voice notifications sent to the first user;

determine a second set of features comprising features related to interaction with the first content when one of the past voice notifications corresponding to the first content was sent;

generate, by a first machine learning model based at least in part on the first data, the first set of features, and the second set of features, a prediction that a voice notification will increase a probability that the first user will interact with the first content;

generating a first activity map comprising a first number of columns and a first number of rows, wherein a first cell at a first intersection of a first column and a first row corresponds to a first time slot during a first week, wherein a value of the first cell represents a first number of occurrences of a first event type during the first time slot and the first week;

generating, by inputting the first activity map into a convolutional neural network comprising at least a first convolutional layer and a first pooling layer, first feature data comprising a down-sampled representation of the first activity map;

generating, by at least one classification layer using the first feature data, a first vector, wherein a first element of the first vector corresponds to a second time slot, and wherein a value of the first element represents a predicted probability that an event of the first event type will occur if the voice notification is sent to a first device during the second time slot;

selecting the second time slot based on the predicted probability associated with the second time slot in the first vector; and

send audio data comprising the voice notification to the first device during the second time slot, wherein the voice notification includes a prompt related to the first content.

14. The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine a first feature representing a subset of the past voice notifications that were consumed by the first user; and

determine a second feature representing a subset of the past voice notifications that were not consumed by the first user, wherein the first set of features comprises the first feature and the second feature.

15. The system of claim 14 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to determine a first feature representing a subset of the past voice notifications that were consumed and which were associated with past interaction with content.

16. The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine a first feature representing orders associated with past voice notifications that were consumed by the first user; and

determine a second feature representing orders associated with past voice notifications that were not consumed by the first user, wherein the second set of features comprise the first feature and the second feature.

17. The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine, by the convolutional neural network, a time at which to send the audio data based at least in part on a timing of a subset of the past voice notifications that were consumed by the first user.

18. The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine, by the convolutional neural network, a time at which to send the audio data based at least in part on historical data indicating one or more times-of-day associated with the first user interacting with a natural language processing system.

19. The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine a third set of features representing a first frequency at which the past voice notifications were sent; and

determine, by a second machine learning model, a second frequency with which to send voice notifications based at least in part on the first set of features, the second set of features, and the third set of features.

20. The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine, by a second machine learning model, the first device from among other devices based at least in part on presence data indicating detection of the first user by a sensor associated with the first device; and

determine, by a third machine learning model a first time of day at which to send the audio data to the first device based at least in part on historical data indicating times at which the first user has historically consumed voice notifications.