- The paper introduces a decision-theoretic framework that quantifies interruption costs and deferral penalties for digital alerts.
- It employs Bayesian inference and SVM-based classification to model user attention and determine email criticality.
- The implementation in PRIORITIES systems demonstrates practical benefits by prioritizing notifications based on the computed Net Expected Value of Alerting.
Attention-Sensitive Alerting: A Technical Overview
The paper "Attention-Sensitive Alerting" by Horvitz, Jacobs, and Hovel introduces utility-directed methods for mediating potentially distracting alerts to computer users. The research focuses on the balance between the context-sensitive costs of deferring alerts and the cost of interruption. The authors discuss the development of models and inference procedures within the context of the Attentional Systems project at Microsoft Research. The primary focus is on the intelligent guidance of email alerts, particularly how to prioritize email by criticality and modulate notifications.
Bayesian Models for User Attention
The researchers employ Bayesian models to infer a probability distribution over a user's attentional focus. By monitoring multiple sources of information, the system assesses the expected utility of messages, inferring a user's attention and the expected cost of interruption. This involves constructing Bayesian models that consider distinct attentional focus states and formulating models that infer a probability distribution over the cost of interruption for different notifications.
Expected Cost of Interruption and Deferral
The utility of relaying alert information is decomposed into expected costs and benefits. The expected cost of an alert's interruption depends on the nature of the interruption and the user's current task focus. The authors introduce a framework for assessing the expected cost of deferring alerts, using the concept of the Expected Cost of Delayed Action (ECDA), which quantifies the cost of not reviewing notifications in a timely manner.
Criticality Classification of Emails
The criticality of email messages is inferred through a machine-learning approach, employing techniques like Support Vector Machines (SVMs) to classify and predict the criticality of incoming email. This classification is integrated into decision-theoretic alerting procedures, allowing the system to compute the Net Expected Value of Alerting (NEVA). The classification is based on features like sender, recipient, time-critical language, and more, enhancing the accuracy of predicting email criticality.
Implementation in PRIORITIES Systems
The research implements these methodologies in the PRIORITIES family of systems, which interface with Microsoft Outlook. These prototypes prioritize email based on criticality and automate actions such as playing sounds or forwarding messages. PRIORITIES-ATTEND is an advanced version that integrates Bayesian networks to make decisions based on inferred user attention and NEVA.
Implications and Future Directions
The paper demonstrates a structured approach to integrating decision-theoretic principles into alert management systems, particularly in the domain of email notifications. The implications extend to broader user-interface systems where managing information flow is crucial. Future research could explore richer models for user attention inference and further integrate these systems with evolving AI and human-computer interaction paradigms.
In summary, this research provides a comprehensive methodological framework for managing digital alerts by balancing interruption costs and alert criticality, paving the way for more refined, context-aware notification systems in the future.