Activity-Centric Email: A Machine Learning Approach

Nicholas Kushmerick, Tessa A. Lau, Mark Dredze, Rinat KhoussainovAAAI 2006

Introduction

While email remains the primary tool for managing professional and personal tasks, current email clients are fundamentally "message-centric." They treat communications as isolated units or simple chronological threads, placing the cognitive burden of "activity management"—tracking the status of complex, multi-step processes like hiring, trip planning, or order fulfillment—entirely on the user. This research proposes a shift toward an "activity-centric" paradigm, utilizing machine learning to automatically group emails into ongoing tasks, map their underlying workflows, and predict future states.

Modeling Structured Activities

The researchers distinguish between two primary types of email activities. The first, Structured Activities, involve rigid and predictable workflows, such as e-commerce transactions. These are modeled as Finite-State Automata (FSA), where states represent the current status of a process and transitions are triggered by incoming emails.

To automate this, the authors employ an unsupervised approach for activity identification, clustering messages using unique identifiers like invoice or tracking numbers. Transition identification is achieved by clustering messages based on their functional role (e.g., "Order Confirmation") using text analysis. Finally, the MDI algorithm is used to induce a regular grammar—effectively a process map—from observed message sequences.

Workflow induction for structured e-commerce activities, showing email examples and the resulting Finite-State Automaton.

Addressing Informal Activities

In contrast, Informal Activities, such as meeting coordination, are less predictable and rely on human negotiation. To manage these, the authors developed specialized similarity metrics:

  • SimOverlap: Compares the social network of participants involved in different messages.
  • SimContent: Utilizes a variant of Latent Semantic Indexing to identify thematic links.

Furthermore, the system employs Speech Act Detection to identify the intent behind a message (e.g., a request or a commitment). By using a collective classification approach, the system iteratively refines its understanding of both the activity grouping and the intent of individual messages, as each provides context for the other.

Interface Prototypes and Evaluation

The proposed methods were evaluated using e-commerce data from vendors like FreshDirect and eBay, as well as internal enterprise workflow data. A key finding was that a hybrid model—combining social context (SimOverlap) and content analysis (SimContent)—outperformed single-metric approaches.

Beyond classification accuracy, the researchers demonstrated the practical utility of these models by integrating them into a prototype "activity-centric" interface for the Thunderbird email client. This interface provides users with an "activities pane" that prioritizes emails based on project status rather than simple arrival time.

Prototypes of activity-centric email interfaces, including an activity inbox and a browser for tracking task status.

Insights and Future Directions

A core insight of this work is that activities are distinct from threads; a single project may span multiple threads, calendar invites, and attachments. The research highlights that social context is often a stronger signal for grouping than keywords alone. Moreover, the ability to "discover" business processes through unsupervised learning allows the system to scale without human-defined templates.

Future work aims to expand this system beyond the email silo, integrating web browsing and calendar data. While structured mail is highly amenable to these techniques, the ambiguity of informal human communication remains a significant challenge. The authors emphasize the need for large-scale user studies to quantify the actual reduction in cognitive load for knowledge workers.

Summary updated

BibTeX
@inproceedings{kushmerick2006activity,
  author = {Nicholas Kushmerick and Tessa A. Lau and Mark Dredze and Rinat Khoussainov},
  title = {{Activity-Centric Email: A Machine Learning Approach}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure