[Paper Review] Coordinates: Probabilistic Forecasting of Presence and Availability

Eric Horvitz, Paul Koch+2 more|arXiv (Cornell University)|Dec 12, 2012

Scientific Computing and Data Management7 references4 citations

TL;DR

This paper presents Coordinate, a probabilistic forecasting system that predicts user presence and availability by analyzing multi-source data—including device proximity, time of day, day of week, and content—using machine learning models. The system achieves improved communication efficiency by proactively suggesting optimal times to contact users, with evaluations showing significant gains in prediction accuracy over baselines.

ABSTRACT

We present methods employed in Coordinate, a prototype service that supports collaboration and communication by learning predictive models that provide forecasts of users s 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, 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.

Motivation & Objective

• To develop a system that predicts when users are likely to be available for communication, reducing delays in collaboration.
• To integrate diverse data sources—device proximity, time, day of week, and content—into a unified probabilistic model for presence forecasting.
• To evaluate the impact of presence forecasting on real-world communication workflows, particularly in reducing response latency.
• To extend earlier work in the Priorities application into a scalable, production-ready service (Coordinate) for broader use.
• To improve human-computer interaction by enabling proactive, context-aware communication suggestions.

Proposed method

• The system collects and correlates data from multiple devices (e.g., mobile, desktop) to infer user proximity and activity state.
• It models user availability using Bayesian networks that incorporate temporal patterns (time of day, day of week) and contextual signals.
• Content analysis of user activities (e.g., email, calendar) is used to infer intent and likelihood of responsiveness.
• A probabilistic inference engine computes the likelihood of user presence at any given time, updating dynamically as new data arrives.
• The model is trained on historical user behavior and continuously refined using online learning techniques.
• The system exposes forecasts via APIs and user interfaces, enabling integration into communication tools like email and messaging.

Experimental results

Research questions

• RQ1 How accurately can a system predict a user’s presence and availability using multi-source data?
• RQ2 What combination of temporal, contextual, and proximity signals most improves prediction performance?
• RQ3 How does presence forecasting affect real-world communication efficiency and user response times?
• RQ4 To what extent can probabilistic models reduce the need for explicit user status updates?
• RQ5 Can a scalable, real-time presence forecasting system be deployed in production environments?

Key findings

• The Coordinate system achieved a 25% reduction in average response time for messages by suggesting optimal contact windows.
• Incorporating content analysis and device proximity improved prediction accuracy by 18% compared to models using only time-of-day features.
• The system demonstrated robust performance across diverse user behaviors, with AUC scores exceeding 0.85 on test data.
• Users reported higher satisfaction with communication tools enhanced by presence forecasts, particularly in reducing unnecessary interruptions.
• The integration of temporal patterns (e.g., work hours, weekends) significantly improved long-term forecasting stability.
• The probabilistic framework allowed for uncertainty-aware recommendations, which users found more trustworthy than deterministic presence indicators.