TL;DR
This paper introduces a novel approach to interruptibility management by incorporating users' social roles (e.g., "work," "private") into AI-driven systems. By understanding these roles, the system can better predict when a user is receptive to interruptions, leading to more personalized and less disruptive information delivery.
What Problem Does This Paper Address?
In our increasingly connected world, devices constantly bombard us with information, leading to frequent interruptions. While current attention management systems try to identify opportune moments for delivery, they often overlook a crucial aspect of human behavior: our intrinsic preferences for being interrupted, which are deeply tied to our social roles and life domains. For instance, being interrupted by a work email while trying to relax at home is far more disruptive than receiving a notification from a close friend. Existing systems lack this nuanced understanding, often offering blunt solutions like "do not disturb" modes that are either too restrictive or fail to account for the complexities of balancing work and personal life. This paper aims to bridge this gap by developing AI systems that can infer and leverage social roles to manage interruptions more effectively.
Methods
The core of this research lies in a two-stage classification approach.
- Social Role Classification: The first stage uses machine learning to classify a user's current social role, primarily distinguishing between "private" and "work" roles. This is achieved by analyzing various features extracted from multi-device usage (smartphones and computers), including application streams, keyboard/mouse interactions, location, and physical activity.
- Interruptibility Classification: In the second stage, the inferred social role from the first stage is fed as an additional feature into an interruptibility classifier. This classifier then predicts the user's interruptibility preference (e.g., "interrupt-private," "interrupt-work").
The researchers explored several popular machine learning classifiers, including Random Forest (RF), which emerged as a strong performer. They also incorporated theories from social science, specifically role theory and boundary management, to inform their feature engineering and model design.
Datasets
The study utilized a custom dataset collected over 5 weeks from 16 participants. This dataset comprised:
- Multi-device sensor data: Information gathered from participants' smartphones (Android) and computers (Windows & macOS). This included application usage, keyboard/mouse activity, location, physical activity, and notification data.
- Experience Sampling Method (ESM) data: Participants were prompted at scheduled intervals or after specific interactions to report their current social role and interruptibility preferences.
This dataset is a novel contribution, specifically designed to capture the interplay between social roles, device usage, and interruptibility preferences in a real-world, multi-device context.
Key Insights and Ideas
- Social roles are critical for interruptibility: The study strongly suggests that understanding a user's social role (work vs. private) significantly improves the accuracy of predicting their interruptibility preferences.
- Within-role interruptibility is dominant: Participants showed a tendency to be more interruptible by events originating from the role they were currently enacting, highlighting a "within-role" interruptibility preference. This supports the idea that people naturally segment their private and work lives.
- Multi-device context matters: Combining data from both smartphones and computers generally leads to better classification performance for social roles and private-related interruptibility compared to using data from a single device type.
- Bias in self-reported roles: The study noted a bias in self-reported social roles, with participants frequently selecting "private" or "work." While this might limit generalization, it also means that even with biased data, the models showed robust performance.
Key Results
- Improved Interruptibility Classification: Incorporating self-reported social roles as features improved the weighted F1-score for interruptibility classification by a statistically significant margin (from M=0.70 to M=0.74 for RF).
- Effective Social Role Classification: The models achieved a median interquartile F1-score of 0.76 to 0.83 for classifying social roles, outperforming the baseline classifier.
- Two-Stage Model Performance: The proposed two-stage social role-based interruptibility classification model showed a statistically significant improvement over traditional models (M=0.73 vs. M=0.70 for RF).
- Multi-Device Gains: Combining phone and computer data significantly improved social role classification for both "private" and "work" roles, and also for "private" interruptibility.
Limitations and Future Work
- Generalizability: The study involved a relatively small sample size (16 participants) and was conducted over 5 weeks. The authors acknowledge that generalizing these results to a broader population is challenging.
- Target Variable Bias: The self-reported social roles were biased towards "private" and "work," which might influence the observed performance and the multi-device t-test results.
- Privacy Concerns: The use of detailed personal data raises significant privacy and ethical concerns, particularly regarding workplace surveillance. The authors suggest socio-technical design approaches and privacy-preserving machine learning techniques.
- Future Directions:
- Finer-grained roles: Exploring more granular social roles (e.g., parent, student, supervisor) to capture more nuanced interruptibility preferences.
- Integration with short-term approaches: Combining social role-based management with existing breakpoint or mental workload-based interruption strategies for a more comprehensive system.
- On-device processing: Implementing privacy-preserving machine learning models that process data directly on the user's device to mitigate privacy risks.
Summary updated