TL;DR
This paper explores how our mood and smartphone usage patterns influence how quickly we respond to notifications. By collecting data from participants over five weeks using smartphones and wearables, they built predictive models that can estimate notification response times, showing that mood and physiological signals significantly improve prediction accuracy. This work is crucial for developing smarter notification systems that respect user well-being and productivity.
What Problem Does This Paper Address?
In our hyper-connected world, smartphones and computers constantly bombard us with notifications. While these are meant to be helpful, they can also be incredibly disruptive, leading to stress, frustration, and decreased productivity. The challenge lies in managing these interruptions effectively. Existing systems often struggle to account for the nuanced factors that influence our receptiveness to notifications, such as our emotional state and current activity. This paper aims to bridge this gap by investigating how individual mood and real-world usage behaviors impact notification response times, with the ultimate goal of creating more intelligent and user-aware notification management systems.
Methods
The researchers conducted an "in-the-wild" study, meaning they collected data from participants in their natural environments over five weeks. They developed two applications, "Balance for Android" and "Balance for desktop," to continuously gather background data on smartphone and desktop usage, including app interactions and notification events. To capture participants' internal states, they employed the Experience Sampling Method (ESM) through questionnaires to assess mood, interruptibility, and social roles. Additionally, participants wore Empatica E4 wristbands to collect physiological signals like Electrodermal Activity (EDA), Blood Volume Pulse (BVP), and Skin Temperature (ST).
For analysis, they pre-processed the collected data to harmonize application names and extract relevant features. They then built participant-specific regression models to predict notification response times. Various regression models were tested, including Linear Regressor, Support Vector Regressor (SVR), Gradient Boosting Regressor (GBR), Random Forest Regressor, and Bayesian Ridge Regressor. Model performance was evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), with nested cross-validation used for hyper-parameter optimization and feature selection.
Datasets
The study collected data from 27 participants over five weeks. This dataset includes:
- 42,270 notifications
- 3,236 ESM responses
- Over 5,920 hours of physiological signals from Empatica E4 wristbands.
The paper highlights this as a particularly heterogeneous and diverse dataset for in-the-wild notification response behavior studies.
Key Insights and Ideas
- Mood Matters: Our emotional state (valence and arousal) significantly influences how quickly we respond to notifications. Lower valence (feeling unhappy/frustrated) and higher arousal (feeling agitated) were associated with longer response times.
- Individual Differences are Key: A one-size-fits-all approach to notification management won't work. Each participant exhibited unique patterns in their response times, influenced by their specific app usage habits and personal characteristics.
- Physiological Signals Add Value: Beyond self-reported mood, physiological data from wearables can provide valuable insights into a user's state and improve the accuracy of predicting response times.
- App Categories Influence Response: Communication apps generate the most notifications, and response times vary significantly across different app categories, suggesting that notification management strategies should be app-aware.
- Context is King: Features like time of day, day of the week, physical activity, and the relationship with the notification sender were found to be important predictors of response time.
Key Results
- The developed regression models achieved better prediction performance than baseline models (mean and median) for most participants.
- The Bayesian Ridge Regressor model, using a combination of mobile features and ESM data, achieved the best overall prediction performance with an average MAE of 0.7764 ms and RMSE of 1.0527 ms.
- Including mood-related features (from ESM and physiological signals) significantly improved the prediction performance of the models.
- Feature importance analysis revealed substantial individual differences in what factors most strongly influenced notification response times for each participant.
Limitations and Future Work
- ESM as Interruption: The ESM questionnaires themselves can act as interruptions, potentially influencing participants' smartphone usage.
- Mood Measurement Reliability: Relying solely on self-reported mood can be influenced by response biases. While physiological data helps, it's not a perfect substitute.
- Data Imbalance and Diversity: The study had limited diversity in age and gender, and the number of participants wearing E4 wristbands was small. Data was also unbalanced across different apps and usage patterns, making it difficult to create a truly generalized model.
- Future Work: The authors suggest increasing participant numbers, especially those using wearables, and exploring clustering participants based on their app usage patterns to develop more tailored prediction models. Further research is also needed to validate these models in real-world scenarios and explore their application in designing intelligent notification management systems.
Summary updated