Quantifying the Persistence of Daily Routines

1. Problem Statement

Daily life is structured by recurring patterns of behavior (sleep, movement, device use) that align biological rhythms with social and occupational demands. While previous research has established that individual behaviors are persistent within specific domains (e.g., communication "social signatures" or mobility patterns), it remains unknown whether this persistence generalizes to holistic daily routines that span multiple behavioral domains simultaneously.

The core research questions are:

• Do individuals exhibit a stable, person-specific distribution of days across a set of recurring "routine types"?
• Is the temporal transition between these routine types also person-specific and stable over time?
• Does this "routine signature" persist across different populations (age, occupation, health status) and observation windows (weeks to months)?

2. Methodology

Data Sources

The study utilized three independent longitudinal datasets comprising 1,086 participants and over 153,000 person-days:

1. Tesserae: 592 information workers in the US (1-year duration).
2. MoMo-Mood: 164 participants in Finland (133 patients with depressive episodes, 31 healthy controls) over 1 year.
3. GLOBEM: 497 university students in the US over 4 academic years (shorter 10-week observation windows per year).

Feature Extraction & Preprocessing

Passive sensing data was aggregated into daily vectors containing 13 features across three domains:

• Sleep: Bedtime, wake time, and duration (derived from wearables or phone lock/unlock status).
• Mobility: Step counts or accelerometer magnitude standard deviation, aggregated into four 6-hour bins (Night, Morning, Afternoon, Evening).
• Device Usage: Screen time duration, aggregated into four 4-hour bins.
• Normalization: Features were z-normalized within each participant to focus on intra-individual variability. Days with missing data were excluded.

Modeling Approach

• Clustering (GMM): The authors employed Gaussian Mixture Models (GMM) to identify latent routine types. Unlike K-Means, GMM allows for soft clustering (probabilistic assignment) and captures feature correlations via covariance matrices.
• Model Selection: The optimal number of clusters ($K$) was determined using the Bayesian Information Criterion (BIC). The analysis consistently identified 8 routine types as sufficient to capture the dominant structure of daily life across all studies.
• Signature Definition:
  • Routine Signature ($\sigma$): A vector representing the proportion of days a participant spends in each of the $K$ routine types, ordered from most to least frequent.
  • Transition Signature: A transition probability matrix ($P \in \mathbb{R}^{K \times K}$) representing the probability of moving from one routine type to another between consecutive days.

Persistence Metrics

To quantify stability and distinctiveness, the authors split each participant's data into two equal time windows and calculated:

• Self-Distance ($d_{self}$): The distance between a participant's signature in the first half vs. the second half (measuring intra-individual stability).
• Reference Distance ($d_{ref}$): The distance between a participant's signature and the signatures of other participants in the same time window (measuring inter-individual distinctiveness).
• Distance Metrics: Primarily Jensen-Shannon Divergence (JSD) for distributional signatures, with Cosine Distance and Bhattacharyya distance used for robustness checks.

3. Key Results

Discovery of Routine Types

• Daily life resolves into a compact set of ~8 interpretable routine types.
• These types emerged unsupervised without calendar data but clearly reflected real-world structures (e.g., standard workdays, weekends/holidays with longer sleep/higher screen use, and nocturnal mobility patterns).
• Dominance: Across all studies, the two most frequent routine types accounted for an average of 57.63% of a participant's days, indicating that individuals spend the majority of their time in just two recurring patterns.

Stability of Routine Signatures

• Intra-individual Stability: A participant's routine signature was significantly more similar to their own past self than to any other individual ($d_{self} < d_{ref}$).
  • Tesserae: $d_{self} = 0.135$ vs. $d_{ref} = 0.164$ ($p < 10^{-3}$).
  • This stability held across both short (30 days) and long (135 days) observation windows.
• Transition Dynamics: The way individuals transition between routine types (the transition matrix) was also person-specific and stable.
  • Tesserae: $d_{self} = 0.481$ vs. $d_{ref} = 0.537$ ($p < 10^{-3}$).
  • This indicates that individuals do not just stick to one routine but cycle through multiple types in a characteristic, non-random order.

Predictors of Persistence

• Demographics: Age and gender showed no significant association with the stability of routine signatures.
• Personality Traits:
  • Conscientiousness: Positively associated with higher signature distances (indicating less stability) in long-window analyses ($p=0.001$).
  • Neuroticism: Positively associated with routine-signature distance in long-window models ($p=0.039$).
  • Note: The positive correlation with distance implies that higher scores in these traits correlate with less rigid/persistent routines in this specific context, though the effect sizes were modest.

4. Key Contributions

1. Holistic Framework: Moves beyond single-domain behavioral analysis (e.g., just sleep or just mobility) to model multi-modal daily routines simultaneously.
2. Routine Signature Concept: Formalizes the "routine signature" as a dual-component metric: (1) time allocation distribution and (2) day-to-day transition dynamics.
3. Validation of Persistence: Demonstrates that these signatures are robust "behavioral fingerprints" that persist across diverse populations (clinical, student, professional) and time scales.
4. Unsupervised Discovery: Proves that a small set of latent routine types (approx. 8) can explain the dominant structure of daily life without prior labeling.

5. Significance and Implications

• Personalized Health Monitoring: Routine signatures can serve as individual-specific baselines. Deviations from a person's own signature (rather than population norms) could detect early signs of health issues, mental health relapses, or lifestyle disruptions.
• Re-identification Risk: The high distinctiveness and temporal stability of these signatures pose a privacy risk. Even in de-identified datasets, combining multiple behavioral modalities could allow for the re-identification of individuals across time and datasets.
• Behavioral Science: Establishes that the "persistence" observed in single behaviors is a general principle of human organization, suggesting that humans naturally settle into a limited set of complex, multi-behavioral patterns to manage cognitive load and environmental demands.

Limitations

• Data Sparsity: Location data (GPS) was excluded due to high missingness, and communication features were sparse (limited to calls/SMS), potentially missing nuances in social interaction.
• Unsupervised Nature: While robust, the clustering is unsupervised; cluster meanings can vary slightly between populations, and the "bag-of-days" approach ignores sequential temporal dependencies (though periodicity like weekday/weekend splits emerged naturally). Future work may incorporate Markov models to better capture temporal sequences.