Attention-Sensitive Alerting

Eric Horvitz, Andy Jacobs, David Hovel

Abstract

As digital connectivity increases, human attention has become a scarce and valuable resource. This paper introduces a decision-theoretic approach to managing interruptions by weighing the "cost of distraction" against the "value of information." The proposed framework utilizes Bayesian models to infer a user's attentional state and Support Vector Machines (SVM) to classify the criticality of incoming messages. By calculating the Net Expected Value of Alerting (NEVA), the system ensures that users are only interrupted when the benefit of immediate information outweighs the cognitive cost of the break.

Introduction

The "Attention Economy" presents a significant challenge in modern computing: while processing power grows exponentially, human cognitive capacity remains fixed. Traditional operating systems and communication tools often treat all alerts with equal urgency, bombarding users regardless of their current task or focus level. This "interruption problem" necessitates a more sophisticated "Attention Manager" capable of reasoning about the costs and benefits of notifications under uncertainty.

Methodology: Inferring Attention and Criticality

The researchers developed a two-pronged probabilistic framework to address the uncertainty of user state and message importance.

1. Modeling Attentional Focus

To estimate a user's current task (e.g., "Deep Work" vs. "Casual Browsing"), the system employs Bayesian Networks. These models synthesize data from various sensors, including desktop activity, ambient acoustics via microphone, active applications, and digital calendars.

Bayesian network for inferring a user's attentional focus from multiple sensory inputs.

2. Assigning Message Criticality

The "benefit" of an alert is evaluated using Support Vector Machines (SVM) trained to classify incoming emails. The classifier identifies features such as the sender's relationship to the user, linguistic cues indicating urgency (e.g., "as soon as possible"), and references to upcoming deadlines.

The system ultimately calculates the Net Expected Value of Alerting (NEVA). This metric determines if the cost of delaying information until a natural break is higher than the immediate cost of interruption.

Evaluation and Implementation

The framework was validated using a corpus of 1,500 emails for training and a test set of 500 emails. Human validation against 200 real-world emails showed a high correlation (0.9) between automated criticality scores and human judgment.

These models were integrated into a prototype suite called PRIORITIES. Working alongside standard email clients like Outlook 2000, PRIORITIES modulates alert modalities—such as adjusting notification sounds, foregrounding windows, or forwarding urgent messages to mobile devices—based on the inferred NEVA.

The PRIORITIES system interface, displaying email messages ranked by their inferred criticality scores.

Key Insights

  • Attention as a Commodity: The research shifts UI design focus from mere "information delivery" to active "attention management."
  • Expected Cost of Delayed Action (ECDA): Criticality is defined by the cost of delay. A message about an imminent meeting is more critical than a high-priority report due in a week.
  • Contextual Meta-Features: The SVM classifier leverages relationship data and recipient lists rather than just keyword matching.

Conclusion and Future Work

While the prototype demonstrated effective discrimination of urgent messages, the authors identified areas for improvement, such as handling "message chunking" to avoid redundant alerts and developing richer models for long-term cognitive load. Future work aims to refine assessment techniques to better quantify the subjective costs of interruption across diverse user tasks.

Summary updated

Abstract

We introduce utility-directed procedures for mediating the flow of potentially distracting alerts and communications to computer users. We present models and inference procedures that balance the context-sensitive costs of deferring alerts with the cost of interruption. We describe the challenge of reasoning about such costs under uncertainty via an analysis of user activity and the content of notifications. After introducing principles of attention-sensitive alerting, we focus on the problem of guiding alerts about email messages. We dwell on the problem of inferring the expected criticality of email and discuss work on the Priorities system, centering on prioritizing email by criticality and modulating the communication of notifications to users about the presence and nature of incoming email.
BibTeX
@inproceedings{horvitz1999attention,
  author = {Eric Horvitz and Andy Jacobs and David Hovel},
  title = {{Attention-Sensitive Alerting}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1999},
  url = {https://openreview.net/forum?id=4M1QkOAQsNV},
}
Figure