Time-preserving embeddings

1. A method, comprising:

executing, on a processor of a computing device, instructions that cause the computing device to perform operations, the operations comprising:

generating user trails of activities performed by users;

identifying frequencies at which the activities were performed;

assigning indices to a set of activities identified from the activities as having frequencies above a threshold, wherein the assigning comprises assigning a first index to a first activity having a first frequency above the threshold and assigning a second index to a second activity having a second frequency above the threshold;

mapping activity descriptions of the set of activities to the indices to generate a vocabulary, wherein the mapping comprises mapping a first activity description to the first index assigned to the first activity and mapping a second activity description to the second index assigned to the second activity; and

training a model using the user trails, timestamps of the activities, and the vocabulary to learn a set of time-preserving embeddings.

2. The method of claim 1 , wherein the training comprises:

generating a set of learned parameters based upon the training of the model.

3. The method of claim 1 , wherein the training comprises:

training the model to generate the set of time-preserving embeddings as user embeddings describing the users.

4. The method of claim 1 , wherein the training comprises:

training the model to generate the set of time-preserving embeddings as activity embeddings describing the activities.

5. The method of claim 1 , wherein the training comprises:

transforming, using unsupervised training, the user trails into fix-length numerical vectors as the set of time-preserving embeddings representing collections of temporally distributed user events.

6. The method of claim 1 , wherein the training comprises:

utilizing a temporal score function to encode temporal information into the set of time-preserving embeddings, wherein the temporal score function is utilized to map a time of event to a score.

7. The method of claim 6 , wherein the temporal score function utilizes a parameterized sigmoid function that outputs the score based upon a sigmoid and one or more trained parameters.

8. The method of claim 1 , comprising:

utilizing the set of time-preserving embeddings to perform a task corresponding to at least one of predicting a user interest, predicting a likelihood a user will perform an action, generating a recommendation to provide the user, or predicting a likelihood that the user will interact with content.

9. The method of claim 1 , comprising:

utilizing a set of learned parameters, derived from the training of the model, to perform a task corresponding to at least one of predicting a user interest, predicting a likelihood a user will perform an action, generating a recommendation to provide the user, or predicting a likelihood that the user will interact with content.

10. A non-transitory machine readable medium having stored thereon processor-executable instructions that when executed cause performance of operations, the operations comprising:

generating user trails of activities performed by users;

generating a vocabulary based upon mappings of activity descriptions to indices, wherein the indices are mapped to a set of activities identified from the activities as having frequencies above a threshold, wherein the mappings comprise (i) a first mapping of a first activity description to a first index assigned to a first activity having a first frequency above the threshold and (ii) a second mapping of a second activity description to a second index assigned to a second activity having a second frequency above the threshold;

training a model using the user trails, timestamps of the activities, and the vocabulary to learn a set of time-preserving embeddings; and

utilizing the set of time-preserving embeddings to perform a first task.

11. The non-transitory machine readable medium of claim 10 , wherein the operations comprise:

utilizing the set of time-preserving embeddings to perform a second task having a task type different than a task type of the first task.

12. The non-transitory machine readable medium of claim 10 , wherein the operations comprise:

generating a set of learned parameters based upon the training of the model; and

utilizing the set of learned parameters to perform the first task.

13. The non-transitory machine readable medium of claim 12 , wherein the operations comprise:

utilizing the set of learned parameters to embed incoming activities.

14. The non-transitory machine readable medium of claim 13 , wherein the operations comprise:

in response to determining that an incoming activity is a trending activity performed a threshold number of times by a threshold number of users, including the trending activity within the vocabulary for a subsequent training of the model.

15. The non-transitory machine readable medium of claim 13 , wherein the operations comprise:

in response to determining that an incoming activity is not within the vocabulary, refraining from adding the incoming activity into the vocabulary until a subsequent update of the vocabulary.

16. The non-transitory machine readable medium of claim 12 , wherein the operations comprise:

utilizing the set of learned parameters to embed a user.

17. The non-transitory machine readable medium of claim 16 , wherein the operations comprise:

embedding activities, of a user trail of the user, that are in the vocabulary; and

considering activities, of the user trail, that are not in the vocabulary during a subsequent update of the vocabulary.

18. A computing device comprising:

a processor; and

memory comprising processor-executable instructions that when executed by the processor cause performance of operations, the operations comprising:

generating a vocabulary based upon mappings of activity descriptions, of activities performed by users, to indices, wherein the indices are mapped to a set of activities identified from the activities as having frequencies above a threshold, wherein the mappings comprise (i) a first mapping of a first activity description to a first index assigned to a first activity having a first frequency above the threshold and (ii) a second mapping of a second activity description to a second index assigned to a second activity having a second frequency above the threshold;

training a model using user trails of the activities, timestamps of the activities, and the vocabulary to learn a set of time-preserving embeddings; and

utilizing the set of time-preserving embeddings to perform a task.

19. The computing device of claim 18 , wherein the set of time-preserving embeddings comprise representations of users, wherein a set of users with activity histories within a similarity threshold are in a neighboring proximity within an embedding space.

20. The computing device of claim 18 , wherein the set of time-preserving embeddings comprise token based URL embeddings.