Classification of messages using learned rules
The subject technology receives, in an application on an electronic device, a message, the message being associated with a user and including information in a header portion of the message. The subject technology determines, on the electronic device, a current state of messaging activity of the user based at least in part on a log of previous events associated with the user, where the log of previous events includes information that has been hashed using a cryptographic hash function. The subject technology determines, on the electronic device using a set of rules provided by a machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user. The subject technology sets, on the electronic device, an indication that the message is important based on the determining.
1. A method comprising:
receiving, in an application on an electronic device, a message, the message being addressed to a user and including information in a header portion of the message;
determining, on the electronic device, a current state of messaging activity of the user based at least in part on a log of previous actions performed by the user, wherein the log of previous events includes information that has been hashed using a cryptographic hash function;
determining, on the electronic device using a set of rules previously provided by a machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user; and
setting, on the electronic device, an indication that the message is important based on the determining.
2. The method of claim 1 , further comprising:
receiving, in the application, a second message, the second message being associated with the user and including second information in a header portion of the second message;
determining, using a particular set of rules, that the second message is important, wherein the particular set of rules are predetermined rules that are provided without utilizing the machine learning model; and
setting an indication that the second message is important based on the determining using the predetermined rules.
3. The method of claim 1 , wherein the set of rules correspond to respective rules that have been learned by the machine learning model prior to receiving the message, and the log of previous actions performed by the user comprises at least an aggregation of actions performed by the user over a period of time.
4. The method of claim 1 , wherein the log of previous actions performed by the user comprises one or more actions performed by the user on messages received by the application.
5. The method of claim 1 , wherein the current state of messaging activity of the user indicates a likelihood that the user will read the message.
6. The method of claim 1 , wherein the current state of messaging activity of the user is based on a number of messages from a sender of the message that the user has not read, replied, sent, or forwarded over a past period of time.
7. The method of claim 1 , wherein the current state of messaging activity of the user is based on a fraction of messages from a sender of the message that the user has replied to over a past period of time.
8. The method of claim 1 , wherein the machine learning model is trained at a server.
9. The method of claim 8 , wherein the machine learning model is trained based at least in part on log data associated with email messages of respective users from respective electronic devices.
10. The method of claim 1 , wherein the log of previous actions performed by the user includes information that has been hashed using a cryptographic hash function and a salt associated with the user, the salt comprises random data, the application comprises an email application, and the message comprises an email message.
11. A system comprising:
a processor;
a memory device containing instructions, which when executed by the processor cause the processor to:
receive, in an application on an electronic device, a message, the message being addressed to a user and including information in a header portion of the message;
determine, on the electronic device, a current state of messaging activity of the user based at least in part on a log of previous actions performed by the user, wherein the log of previous events includes information that has been hashed using a cryptographic hash function;
determine, on the electronic device using a set of rules previously provided by a machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user; and
set, on the electronic device, an indication that the message is important based on the determining.
12. The system of claim 11 , wherein the memory device contains further instructions, which when executed by the processor further cause the processor to:
receive, in the application, a second message, the second message being associated with the user and including second information in a header portion of the second message;
determine, using a particular set of rules, that the second message is important, wherein the particular set of rules are predetermined rules that are provided without utilizing the machine learning model; and
set an indication that the second message is important based on the determining using the predetermined rules.
13. The system of claim 11 , wherein the set of rules correspond to respective rules that have been learned by the machine learning model prior to receipt of the message, and the log of previous actions performed by the user comprises at least an aggregation of actions performed by the user over a period of time.
14. The system of claim 11 , wherein the log of previous actions performed by the user comprises one or more actions performed by the user on messages received by the application.
15. The system of claim 11 , wherein the current state of messaging activity of the user indicates a likelihood that the user will read the message.
16. The system of claim 11 , wherein the current state of messaging activity of the user is based on a number of messages from a sender of the message that the user has not read, replied, sent, or forwarded over a past period of time.
17. The system of claim 11 , wherein the current state of messaging activity of the user is based on a fraction of messages from a sender of the message that the user has replied to over a past period of time.
18. The system of claim 11 , wherein the current state of messaging activity of the user is based on a number of times the user has flagged a respective message from a sender of the message over a past period of time.
19. The system of claim 11 , wherein the log of previous actions performed by the user_includes information that has been hashed using a cryptographic hash function and a salt associated with the user.
20. A non-transitory computer-readable medium comprising instructions, which when executed by a computing device, cause the computing device to perform operations comprising:
receiving, in an application on an electronic device, a message, the message being addressed to a user and including first information in a header portion of the message;
determining, on the electronic device, a current state of messaging activity of the user based at least in part on a log of previous actions performed by the user, wherein the log of previous events includes second information that has been hashed using a cryptographic hash function, the second information that has been hashed corresponding to data only from a header portion of a message and excluding data from a body portion of the message;
determining, on the electronic device using a set of rules previously provided by a machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user; and
setting, on the electronic device, an indication that the message is important based on the determining.
21. The method of claim 1 , wherein determining, on the electronic device using the set of rules previously provided by the machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user comprises determining, on the electronic device using the set of rules previously provided by the machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user and the information in the header portion of the message.
22. The method of claim 21 , wherein the information that has been hashed using the cryptographic hash function comprises information associated with one or more other users of one or more other electronic devices, and wherein the information that has been hashed using the cryptographic hash function has been generated in part by performing privacy performing operations for the one or more other users, the privacy preserving operations including:
hashing only a subset of the information associated with the one or more other users, the subset corresponding to data only from header portions of one or more prior messages received at the electronic device in association with the one or more other users and excluding data from body portions of the one or more prior messages; and
anonymizing the header portions of the one or more prior messages by hashing the header portions of the one or more prior messages using a salt.
23. The method of claim 22 , wherein the determining, on the electronic device using the set of rules previously provided by the machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user and the information in the header portion of the message comprises determining, on the electronic device using the set of rules previously provided by the machine learning model, that the user is likely to view the message based on the current state of the messaging activity of the user and the information in the header portion of the message after attempting to determine whether the message is important to the user using a curated rule that is provided without utilizing the machine learning model.