TL;DR
This study explores how mobile phone usage reflects emotional states and social dynamics during home confinement, using the COVID-19 lockdown as a case study. They found that app usage patterns can effectively predict emotional well-being, offering a path towards developing digital interventions for individuals in various confinement scenarios.
What Problem Does This Paper Address?
Home confinement, whether due to illness, disability, or remote work, significantly impacts mental well-being and digital device usage. Traditional methods for assessing well-being (surveys) are time-consuming and prone to bias. While mobile sensing offers a more objective approach, existing methods often struggle with the limited physical movement during indoor confinement. This paper addresses this gap by investigating how granular app usage data can provide deeper insights into emotional well-being and social dynamics during confinement, and whether this data can be used to predict emotional states.
Methods
The researchers developed a novel unsupervised approach called App-based Information Gain Temporal Segmentation (AIGTS) to identify distinct user activities from app usage data. AIGTS segments app usage sequences and then clusters these segments to identify common behavioral patterns. They also employed the Experience Sampling Method (ESM) to collect self-reported data on emotional states, social roles, and work-life balance. For prediction, they used regression models (KNN, Random Forest, Gradient Boosting) to forecast valence and arousal, employing feature selection based on the F-statistic and nested cross-validation for robust evaluation.
Datasets
The study collected data from 32 participants in Melbourne, Australia, during a three-week COVID-19 lockdown period. This dataset includes:
- 502,485 records of smartphone and desktop usage.
- 1,749 ESM responses.
- 265 end-of-day surveys.
This dataset is described as a unique, real-world snapshot of emotional well-being and social dynamics during lockdown.
Key Insights and Ideas
- App usage as a window into emotions: The study demonstrates that app usage patterns are strong indicators of emotional states, offering a non-intrusive way to understand well-being during confinement.
- Time is a critical factor: Features related to the time of day and day of the week were consistently the most important predictors of valence and arousal, highlighting the significance of temporal context in understanding emotional states.
- Group-specific models matter: Prediction performance varied significantly across different user groups (e.g., early vs. late workers), suggesting that personalized or group-specific models are crucial for accurate emotional prediction.
- Work-life balance challenges: Participants reported longer work hours and blurred boundaries between personal and professional life, underscoring the impact of confinement on work-life balance.
- Social roles influence emotions: Engaging in private-related roles was associated with higher valence and lower arousal, suggesting that personal activities can foster well-being during confinement.
Key Results
- Valence and Arousal Prediction: The Random Forest (RF) regressor demonstrated superior performance in predicting valence and arousal compared to baselines, achieving significant reductions in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). For instance, the RF regressor achieved a 17.23% reduction in RMSE for valence compared to the Median baseline.
- Feature Importance: Time-related features (e.g.,
sin_time,cos_time) and app category features were identified as the most important predictors for both valence and arousal. - Group Model Performance: For late workers, RF models showed improvements in prediction accuracy compared to baselines, while for early workers, the models performed similarly to baselines, suggesting that factors beyond app usage might influence their emotions.
- Individual Differences: The study observed significant variability in prediction performance across individual participants, highlighting the need for personalized models.
Limitations and Future Work
- Time and Geographical Constraints: The study was conducted in Melbourne during a specific lockdown period, limiting the generalizability of findings to other regions with different restrictions.
- Short Duration of Data Collection: The three-week data collection period might not fully capture the long-term evolution of well-being and digital device usage during prolonged confinement.
- Device Usage and Untracked Data: The study acknowledges the potential for incomplete data due to participants using untracked devices, which could have influenced emotional states and introduced noise.
- Sample Representativeness: The sample, recruited via social media, may be skewed towards tech-savvy individuals and information workers, potentially limiting generalizability to the broader population.
Future work includes expanding the research to investigate desktop usage behavior and comparing it with smartphone usage to gain a deeper understanding of daily lives and work-life balance. The authors also suggest the need for longer data collection windows and more extensive device tracking in future studies.
Summary updated