ABSTRACT:
A previous study shows that busy professionals receive in excess of 50 emails per day of which approximately 23% require immediate attention, 13% require attention later and 64% re unimportant and typically ignored. The flood of emails impact mobile users even more heavily.
Flooded inboxes cause busy professionals to spend considerable amounts of time searching for important messages, and there has been much research into automating the process using email content for classification; but we find email priority depends also on user context. In this paper we describe the Personal Messaging Assistant (PMA), an advanced rule-based email management system which considers user context and email content.
Context information is gathered from various sources including mobile phones, indoor and outdoor locationing systems, and calendars. PMA uses separate scales of importance and urgency to prioritize emails and to decide on an appropriate action, such as SMS to user, defer to later, file or forward.
Initial results yield 96% recall and 88% precision in importance classification of emails; 95% recall and 92% precision in urgency classification of emails. PMA shows a 30X reduction in false-negative rates over existing systems. A key contribution of our work has been to leverage an extensible set of context information, gathered in a mobile environment, for the classification of emails and customizable decision making.
SNIPPETS:
Approach – Ranking Emails:
The primary objective of the Personal Messaging Assistant (PMA) is to effectively prioritize and categorize emails for immediate or delayed delivery, forwarding or filing, by taking context information into consideration. The system decides to deliver specific emails to the user as and when an email (i.e., its content and envelope) becomes important and/or urgent in the current context. It also decides on the appropriate mode of delivery (e.g., SMS, Text-to-Voice on the mobile phone, IM/XMPP) based on the current context of the user.
Classifications of Importance and Urgency.
Context Representation:
Context is the differentiating factor that makes PM. A unique in the way it handles email. We classify context information into two categories: static context and dynamic context. As their names suggest, static context encapsulates the user preferences, and priorities that either remain relatively constant over time or change gradually
Types of Context.
Architecture:
The PMA architecture consists of five main components, shown in Fig.3. They are the email preprocessor, the context-generator, the importance processor, the urgency processor and the delivery agent.
Architectural Overview of the PMA System.
Content representation of java class.
TESTING AND RESULTS:
Baseline:
Initially statistical data from several user mailboxes was collected and analyzed to observe how many emails were received each day, and the number of emails that were left unread. Emails in the main inbox were counted separately from emails filed manually or automatically filed in folders or the Trash folder.
The statistical data regarding the amount of emails deleted by users could not be accurately discovered, since the Trash folder is periodically cleared by the email system. Therefore the baseline is an underestimate of the actual complexity. Data was collected for a 120 day period except for Trash folder data which was collected for a 30 day period.
Summary per-day Baseline Email Statistics.
PMA Action Sampling:
The third stage of testing was to evaluate the percentages of various actions taken on emails (i.e., SMSed to the user, filed for later viewing, forwarded to a peer) and the effectiveness of these actions in different context situations.
Variation of Actions Taken on Emails with Change in Context.
FUTURE WORK:
Performance, generalization and personalization are three dimensions that could be improved in the future versions of the system, which is the major focus for our future work. The performance of the system is two dimensional. It can be improved in terms of scanning email content and learning the user context information.
Advanced machine learning schemes could be used to automate the learning of keywords from user feedback; Naïve-Bayesian methods are under consideration for this purpose. Also, during the initial iterations of the development of PMA, the system run-time was not a major concern, as more emphasis was placed on using context information for email sorting. Going forward, system performance could be improved to consume less system time on larger mail boxes.
Lexical analysis or thesaurus based canonic word form analysis on email content to replace the word stemming method used in the current prototype is under consideration. The next steps to generalize the PMA are to make the system work with other email client accounts like Yahoo! mail and Hotmail and adding handling support for additional message types like SMS, IM, RSS, HTTP and Voice.
Personalization of the system is another area that would be addressed in future research. This includes the creation of a user interface to allow users to create/edit custom rules. Also planned is a user interface on the mobile device to allow user feedback regarding actions taken on emails.
Determining the costs of data transfer and processing on the mobile device in terms of power and bandwidth are areas of research under consideration for future work. We also plan to conduct a larger scale usability test to study the effectiveness of the PMA system in comparison to human-based sorting. The study will also attempt to gather users’ requirements for customization.
CONCLUSION:
The key goal of PMA is to make email processing for users not just easy but more accurate. PMA considerably reduces time spent by users on filtering emails by sorting and delivering messages that are relevant to the user in his current context.
Unlike other email filtering systems that depend solely on email content for filtering and sorting, PMA takes into account the content of emails and the contextual information of the user. Another unique aspect of the system is the consideration of urgency and importance of an email as separate dimensions for classification; thereby PMA is able to integrate a temporal aspect into email sorting.
By combining the use of context-information and email content in classification with the idea of separate scales for importance and urgency, PMA is able to intelligently decide whether an email is to be delivered immediately via SMS, deferred for later delivery forwarded to another address, filed, etc. It is scalable for all inbox sizes and types and offers better performance in terms of identifying email priorities. It could be easily personalized to suit the requirements of any user for better accuracy.
The system is highly efficient over continuous usage compared to discontinuous usage. This system saves ample time for users struggling with email flooding, by filtering emails and decision making on behalf of the user. It gives better performance in terms of filtering emails compared to existing rule based systems. It goes beyond simple email filtering to integrate context-awareness, in an extendable manner, to perform advanced email management.
Source: Carnegie Mellon Silicon Valley
Authors: Senaka Buthpitiya | Deepthi Madamanchi | Sumalatha Kommaraju | Martin Griss