Abstract
We present methods employed in Coordinate, a prototype service that supports collaboration and communication by learning predictive models that provide forecasts of users' presence and availability. We describe how data is collected about user activity and proximity from multiple devices, in addition to analysis of the content of users' calendars, the time of day, and day of week. We review applications of presence forecasting embedded in the Priorities application and then present details of the Coordinate service that was informed by the earlier efforts.
Introduction
Effective collaboration often relies on "opportunistic communication"—the ability to catch a colleague at the right moment. While digital calendars provide some insight, they are frequently inaccurate; users may skip meetings, stay late at lunch, or appear "busy" while remaining open to urgent interruptions. Most existing tools only display a user's current state (e.g., "Away for 5 minutes"). Coordinate addresses this by reasoning about uncertainty and the future, answering critical questions like: "When is this person likely to return?" or "What is the probability they will attend their next scheduled meeting?"
System Architecture
Coordinate is a server-based service built on the lessons learned from an earlier project called Priorities. It is designed to fuse multi-modal data into actionable forecasts through a pipeline of data acquisition, coalescence, and inference.
- Data Acquisition: The system monitors keyboard/mouse activity, ambient acoustics (to detect conversation), and Outlook calendar data across multiple devices. It also incorporates location-based clues like wireless signal strength and GPS.
- Data Coalescence: Multi-device events are aggregated into an XML-encoded database.
- Case Extraction: Rather than a single static model, Coordinate identifies a "reference class" of historical cases similar to the current query (e.g., "Monday mornings when the user has been away for 20 minutes").
- Learning and Inference: Using the WinMine toolkit, the system performs a Bayesian structure search to automatically construct a network tailored to the specific prediction task.
A key innovation is the Expected Cost of Interruption (ECI). This metric calculates a "dollar-value" cost of disturbing a user by combining the probability of meeting attendance with the predicted interruptibility level of that specific event.
Predictive Insights
The system's effectiveness stems from several core observations about user behavior:
- Proximal Activity Context: The time already elapsed since a user changed state is a powerful predictor of future transitions. For instance, the probability of a user returning within a specific window changes dynamically as their "time away" increases.
- Calendars as Noisy Evidence: Coordinate treats calendar entries as probabilistic evidence rather than ground truth. It learns to predict attendance based on features like the meeting organizer, the number of invitees, and the presence of the user's direct reports.
- Dynamic Reference Classes: To mitigate data sparseness, the system uses "heuristic sequencing," starting with specific contexts and progressively broadening search criteria until enough historical data is found for a reliable prediction.
Evaluation and Applications
In evaluations using six months of annotated calendar data, the Bayesian model achieved 0.92 accuracy in predicting meeting attendance and 0.81 accuracy in assigning correct interruptibility levels (low, medium, or high).
These models power several practical applications:
- TimeWave: Automatically updates a user's "Free/Busy" status on a shared calendar with a forecast (e.g., "Likely back in 6 minutes").
- SmartOOF: An intelligent out-of-office agent that informs senders when a recipient is likely to see their message based on the communication's urgency.
Conclusion and Future Work
While Coordinate demonstrates the power of probabilistic forecasting, challenges remain. The current system requires a degree of manual data labeling for training, which may hinder widespread adoption. Future work involves scaling the system beyond small research teams and addressing the significant privacy implications of sharing detailed availability forecasts across an organization.
Summary updated