Multi-tiered approach to E-mail prioritization

• US 9,256,862 B2
• Filed: 06/20/2012
• Issued: 02/09/2016
• Est. Priority Date: 02/10/2012
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Lai, Jennifer C., Lu, Jie, Pan, Shimei, Wen, Zhen
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): Smith, Paulinho E
• Application Number:
  US13/528,598
• Publication Number:
  US 20130339276A1
• Time in Patent Office: 1,329 Days
• US Class Current: 1/1
• CPC Class Codes: G06Q 10/107 Computer-aided management of electronic mailing [e-mailing]

Abstract

A method of automating incoming message prioritization. The method including training a global classifier of a computer system using training data. Dynamically training a user-specific classifier of the computer system based on a plurality of feedback instances. Inferring a topic of the incoming message received by the computer system based on a topic-based user model. Computing a plurality of contextual features of the incoming message. Determining a priority classification strategy for assigning a priority level to the incoming message based on the computed contextual features of the incoming message and a weighted combination of the global classifier and the user specific classifier. Classifying the incoming message based on the priority classification strategy.

Claims

1. A method of automating incoming message prioritization, the method comprising:

training a global classifier using message-level contextual features computed from a plurality of e-mail messages and a priority level assigned to each of the plurality of e-mail messages;

dynamically training a user-specific classifier using message-level contextual features computed from a plurality of feedback instances provided by a user regarding a priority level of previous incoming e-mail messages to the user;

dynamically assessing a message-specific quality of the user-specific classifier by computing a vector similarity or distance between the vector of message-level contextual features of an incoming message against the vectors of message-level contextual features of the plurality of feedback instances provided by the user;

selecting a priority classification strategy from a plurality of priority classification strategies based on the assessed quality of the user-specific classifier, the priority classification strategy using at least one of the global classifier and the user-specific classifier; and

classifying the incoming message based on the selected priority classification strategy.

2. The method according to claim 1, wherein the plurality of priority classification strategies comprises a dynamic linear combination scheme with instance matching based on comparing the vector of the message-level contextual features of the incoming message and the vectors of the message-level contextual features of the plurality of feedback instances, the dynamic linear combination scheme with instance matching comprising:

assessing a quality of the user-specific classifier; and

assigning a weight to each of the global classifier and the user-specific classifier for a linear combination thereof, based on the assessed quality of the user-specific classifier.

3. The method according to claim 1, wherein the plurality of priority classification strategies comprises a dynamic linear combination scheme with instance matching, and

wherein, when the incoming message and a feedback instance of the plurality of feedback instances have at least one of a same sender and subject, the dynamic linear combination scheme with instance matching assigns a same priority to the incoming message as a priority assigned to the feedback instance having at least one of the same sender and subject.

4. The method according to claim 3, wherein, when the incoming message does not have at least one of the same sender and subject as any of the plurality of feedback instances, the dynamic linear combination scheme with instance matching assigns a weight to each of the global classifier and the user-specific classifier for a linear combination thereof.

5. The method according to claim 1, further comprising:

inferring a topic of the incoming message received by the computer system based on a topic model created from the interaction history between the user and the sender of this incoming message; and

computing a message-level contextual feature of the incoming message based on the inferred topic of the incoming message.

6. The method according to claim 5, further comprising:

calculating a first percentage of previously received messages that have a substantially similar topic as the inferred topic of the incoming message;

calculating a second percentage of the previously received messages that have the substantially similar topic which are determined to have been read;

calculating a third percentage of the previously received messages that have the substantially similar topic which are determined to have been at least one of forwarded, replied, saved, and flagged; and

computing a contextual feature of the incoming message by dynamically combining the first percentage, the second percentage, and the third percentage.

7. The method according to claim 5, further comprising computing a plurality of message-level contextual features of the incoming message based on the inferred topic of the incoming message.

8. The method according to claim 5, further comprising computing a message centric feature of the plurality of message-level contextual features based on a percentage of received messages comprising a substantially similar topic as the inferred topic of the incoming message.

9. The method according to claim 1, wherein the plurality of priority classification strategies comprises a dynamic linear combination scheme with instance matching.

10. The method according to claim 9, wherein the dynamic linear combination scheme with instance matching includes assigning a weight to each of the global classifier and the user-specific classifier for a linear combination thereof, based on an assessed quality of the user-specific classifier.

11. The method according to claim 1, wherein the method is performed in an apparatus including an input to receive an incoming message, a processor, and a memory tangibly embodying a set of instructions executed by the processor to perform the automating of a prioritization of the incoming message.

12. A method of automating a prioritization of an incoming message, the method comprising:

creating a plurality of topic models for a user, each topic model to encode an interaction history that the user has with one of the user'"'"'s e-mail contacts, and relationship data with the user and one of the user'"'"'s e-mail contacts;

computing a plurality of message-level contextual features of a plurality of e-mail messages received by the user, based on a content of the messages and the interaction history, the topic models, and the relationship data;

training a global classifier using the plurality of message-level contextual features computed from the plurality of e-mail messages and a priority level assigned to each of the plurality of e-mail messages;

dynamically training a user-specific classifier with a plurality of feedback instances provided by a user regarding a priority level of previous incoming e-mail messages to the user;

dynamically assessing a message-specific quality of the user-specific classifier by comparing the vector of the message-level contextual features of an incoming message against the vectors of the message-level contextual features of the plurality of feedback instances provided by the user;

selecting a priority classification strategy from a plurality of priority classification strategies based on the assessed quality of the user-specific classifier, the priority classification strategy using at least one of the global classifier and the user-specific classifier; and

classifying the incoming message based on the selected priority classification strategy.

13. A non-transitory tangible computer-readable medium embodying a program of machine-readable instructions executable by a digital processing apparatus to perform an instruction control method of automating a prioritization of an incoming message, the instruction control method comprising:

training a global classifier using message-level contextual features computed from a plurality of e-mail messages and a priority level assigned to each of the plurality of e-mail messages;

dynamically training a user-specific classifier using message-level contextual features computed from a plurality of feedback instances provided by a user regarding a priority level of previous incoming e-mail messages to the user;

dynamically assessing a message-specific quality of the user-specific classifier by computing a vector similarity or distance between the vector of message-level contextual features of an incoming message against the vectors of message-level contextual features of the plurality of feedback instances provided by the user;

selecting a priority classification strategy from a plurality of priority classification strategies based on the assessed quality of the user-specific classifier, the priority classification strategy using at least one of the global classifier and the user-specific classifier; and

classifying the incoming message based on the selected priority classification strategy.

14. The non-transitory tangible computer-readable medium according to claim 13, wherein, when the similarity or distance is above or below a predetermined threshold, the user-specific classifier is assigned a weight that is greater than a weight assigned to the global classifier.