Sensing Noticeability in Ambient Information Environments

Research Background and Problem

• Problem or Challenge: Designing notifications in augmented reality (AR) environments that are perceivable yet non-intrusive is a critical but complex issue. Most existing AR systems neglect user context and fail to automatically detect whether users notice notifications. This limitation restricts the system's ability to seamlessly deliver notifications during user activities.
• Significance: Notifications that are overly prominent can disrupt users' primary tasks, while overly subtle designs may be entirely overlooked. Addressing this issue can enhance the intelligence and usability of AR systems.
• Research Motivation and Related Work: This study aims to predict whether users detect notifications in peripheral environments using sensor data. A literature review indicates that while notification research for screens and mobile devices is relatively mature, the AR domain lacks studies on dynamically adjusting notifications and validating their effectiveness.

Solution

• Method or Solution: The authors propose a sensor-based classifier to predict whether users detect peripheral notifications during productive tasks.
  • Collect various sensor data, including user gaze, head position, device interactions, and self-reported engagement levels.
  • Develop machine learning models to analyze these data.
• Innovations:
  • Conducted the first long-term data collection study (98 hours, 12 participants) in a near-realistic environment to predict notification perceivability.
  • Used sensor data to detect implicit user behaviors rather than relying on explicit feedback.
  • Proposed a model combining gaze and engagement data, which outperformed existing methods in accuracy.
• Implementation Steps:
  • Set up an office environment with ambient displays.
  • Randomly display notifications every 5 to 10 minutes and collect user feedback 30 to 60 seconds later.
  • Extract features such as gaze, head posture, and device interactions to train and evaluate various machine learning models.
  • Compare model performance based on different time windows, feature combinations, and personalized versus general models.

Research Outcomes

• Specific Results:
  • The classifier achieved an AUC of 0.81 when combining gaze and self-reported engagement data, and an AUC of 0.76 using gaze data alone.
  • Outperformed existing notification perceivability models, achieving reasonable performance even with limited user data.
  • Models using features centered around the notification time window demonstrated the best predictive performance, indicating that user behavior during the notification period is most informative.
• Advantages:
  • Enables detection of notification perceivability through implicit behaviors without requiring explicit user responses.
  • Provides a method starting from a general model without needing extensive personalized data.
  • Opens up possibilities for gradually increasing notification salience, making notifications more "noticeable but non-intrusive."
• Experimental or Evaluation Results:
  • The primary model (based on gradient-boosted classifiers) performed well across different scenarios and user combinations.
  • Personalized models did not significantly outperform general models but were useful in capturing individual differences in certain cases.
• Limitations and Future Directions:
  • The study is limited to simulated environments, specific task types, and a single notification design (transparency changes).
  • Unable to capture user behavior in fully "wild" conditions; future work should extend validation to more natural settings.
  • Current engagement features rely on self-reports or proxy measurements (e.g., EEG), requiring development of more convenient capture methods.
  • Notification type and timing patterns did not fully simulate real-world environments; future studies should incorporate more natural variables.

Future work in this field can advance notification interaction systems by expanding experiments to more diverse scenarios, notification design types, and multi-device setups. Researchers should also carefully balance considerations of user privacy and immersive experience.