TL;DR
MailCat is an intelligent assistant that helps users organize their email by predicting the three most likely folders for incoming messages. It uses a text classifier trained on a user's existing mail habits, significantly reducing the effort required to file emails and encouraging better organization.
What Problem Does This Paper Address?
Email overload is a persistent problem, and while most email clients offer folder organization for retrieval, the act of filing itself can be a significant cognitive burden. Users often fall behind, leading to cluttered inboxes. This paper addresses the challenge of making email organization less time-consuming and more intuitive, aiming to encourage users to actively manage their inboxes.
Methods
MailCat employs a text classifier to learn a user's email filing habits. When a new email arrives, MailCat predicts the three folders the user is most likely to file it in and presents these as shortcut buttons. The core innovation lies in its adaptive learning approach:
- Initial Training: Upon installation, MailCat analyzes the user's existing folders and previously filed messages to build an initial TF-IDF text classifier. This allows it to start making predictions immediately.
- Incremental Learning: MailCat continuously updates its classifier as the user interacts with their email. This is crucial because user habits and folder contents can change over time. The system can quickly adapt to new folders or shifts in message types within existing folders.
- Shortcut Buttons: Instead of automating filing entirely, MailCat offers three predicted folders as shortcuts. This approach balances automation with user control, reducing the cognitive load without forcing a single, potentially incorrect, automated action.
Datasets
The paper describes experiments conducted using the mailboxes of six real users. No new benchmark datasets were introduced; instead, the system was evaluated on existing, personal email archives.
Key Insights and Ideas
- The "Filing Barrier": The paper highlights that the cognitive effort and interface interaction required for filing emails, not just the number of folders, is a significant barrier to effective organization.
- Three is the Magic Number: Providing three predicted folders, rather than just one, dramatically reduces MailCat's error rate (by half). This suggests a sweet spot for predictive assistance that balances accuracy with user choice.
- Leveraging Existing Data: Bootstrapping the classifier with a user's existing mail archive is a practical and effective way to quickly train an intelligent assistant, avoiding the "cold start" problem.
- Adaptability is Key: For an email assistant to remain useful, it must continuously adapt to evolving user behavior and folder structures. Incremental learning is essential for this.
Key Results
MailCat demonstrated significant accuracy in predicting user filing habits:
- With a single button, MailCat achieved accuracies between 59.8% and 81.1%.
- By providing three shortcut buttons, MailCat's accuracy improved to between 80.2% and 98.1%.
- This means that using three buttons effectively halved the error rate compared to using just one.
- The authors state that MailCat's prediction accuracy is generally between 80% and 90%, even for users with up to 60 folders.
Limitations and Future Work
- Limitations: The paper doesn't explicitly detail limitations beyond the inherent challenge of classification accuracy. The success of MailCat is dependent on the user having a sufficient history of filed emails for initial training.
- Future Work: The authors suggest that the core concepts behind MailCat can be applied to organizing other types of electronic documents, such as bookmarks, files, and audio recordings, within hierarchical folder structures.
Summary updated