PaperlayerView exact revision on arXiv

Source-linked AI summary

DiversityOne: A Multi-Country Smartphone Sensor Dataset for Everyday Life Behavior Modeling

Matteo Busso, Andrea Bontempelli, Leonardo Javier Malcotti, Lakmal Meegahapola, Peter Kun, Shyam Diwakar, Chaitanya Nutakki, Marcelo Dario Rodas Britez, Hao Xu, Donglei Song, Salvador Ruiz Correa, Andrea-Rebeca Mendoza-Lara, George Gaskell, Sally Stares, Miriam Bidoglia, Amarsanaa Ganbold, Altangerel Chagnaa, Luca Cernuzzi, Alethia Hume, Ronald Chenu-Abente, Roy Alia Asiku, Ivan Kayongo, Daniel Gatica-Perez, Amalia de Götzen, Ivano Bison, Fausto Giunchiglia

arXiv:2502.03347v1cs.CYcs.SI

TL;DR

Existing smartphone-sensing datasets often provide limited evidence across countries, constraining study of behavioral variation and model generalization. DiversityOne brings together smartphone sensor data and self-reports from students across eight countries. The dataset supports research on behavior modeling, including cross-country generalization and adaptation.

  • Problem

    Many existing datasets focus on limited populations, sensor sets, or regions, leaving cross-country behavioral variation and model generalization insufficiently represented.

    Source: §1 · p. 2, §2.2 · p. 5

  • Method

    DiversityOne combines smartphone sensor data with questionnaires and self-reports collected through a cross-country study design.

    Source: §1 · pp. 3–4, §3 · pp. 10–11

  • Results

    DiversityOne provides a multi-country dataset for studying user behavior and developing models for domain generalization, adaptation, transfer learning, and self-supervised learning.

    Source: §1 · pp. 3–4

  • Takeaways & Limitations

    The dataset offers a resource for investigating behavior modeling and machine-learning approaches across countries.

    Source: §1 · p. 3

    Participation in the intensive longitudinal survey varied by country, with some AMRITA participants completing below 40%.

    Source: §6.3 · p. 31

Abstract

from arXiv · show

Understanding everyday life behavior of young adults through personal devices, e.g., smartphones and smartwatches, is key for various applications, from enhancing the user experience in mobile apps to enabling appropriate interventions in digital health apps. Towards this goal, previous studies have relied on datasets combining passive sensor data with human-provided annotations or self-reports. However, many existing datasets are limited in scope, often focusing on specific countries primarily in the Global North, involving a small number of participants, or using a limited range of pre-processed sensors. These limitations restrict the ability to capture cross-country variations of human behavior, including the possibility of studying model generalization, and robustness. To address this gap, we introduce DiversityOne, a dataset which spans eight countries (China, Denmark, India, Italy, Mexico, Mongolia, Paraguay, and the United Kingdom) and includes data from 782 college students over four weeks. DiversityOne contains data from 26 smartphone sensor modalities and 350K+ self-reports. As of today, it is one of the largest and most diverse publicly available datasets, while featuring extensive demographic and psychosocial survey data. DiversityOne opens the possibility of studying important research problems in ubiquitous computing, particularly in domain adaptation and generalization across countries, all research areas so far largely underexplored because of the lack of adequate datasets.

1 Introduction

DiversityOne addresses the lack of datasets capturing behavioral diversity across countries by combining smartphone sensing with self-reports from students in eight countries. Its scale and breadth support research on cross-country behavior modeling and model generalization.

Motivation

  • Existing datasets often center on Global North populations, limiting evidence about how cultural and geographic variation relates to everyday behavior and model generalization.Differences in eating, sleep, and social routines can affect smartphone use and sensor-based inferences.Source: §1 · pp. 2–3

Dataset

  • DiversityOne includes 782 college students across eight countries, with four weeks of raw data from 26 smartphone sensors and repeated in-situ self-reports.Reports cover activities, locations, social contexts, moods, sleep quality, and daily expectations; sensor streams are organized into six bundles.Source: §1 · p. 3

Contributions

  • The dataset combines questionnaires, passive smartphone sensing, and self-reports, supporting behavioral inference and research on domain generalization, adaptation, transfer learning, and self-supervised learning.The study received institutional ethical approvals and follows GDPR requirements.Source: §1 · p. 3

Paper contributions

  • The authors describe a cross-country study design, dataset-quality validation, access resources, and initial insights and recommendations for future research.Their recommendations address study design and cross-country collection as well as multi-country sensor-data analysis.Source: §1 · p. 4

2 Background and Related Work

The background motivates combining diverse smartphone sensors with contextual self-reports and a cross-country perspective. It frames DiversityOne through social practices and contrasts its breadth with existing datasets.

2.1 Smartphone Sensing for Behavior Modeling

  • Smartphone sensing uses passive sensor streams and interaction data to infer behaviors ranging from activity and social context to mental states, with models applied to unseen data.The literature includes both traditional machine-learning methods and deep-learning techniques.Source: §2.1 · p. 4

2.2 Continuous and Interaction Sensing Modalities

  • Existing public datasets often lack raw data and depth across continuous and interaction sensing, motivating their integration for more detailed behavioral analysis.Continuous sensing captures passive signals, while interaction sensing records user engagement with devices.Source: §2.2 · pp. 4–5

2.3 Effect of Country Diversity on Sensing and Self-reported Ground Truth

  • Country differences can shape sensor patterns and self-reported labels, while many prior studies overlook these differences, challenging cross-country model generalization.International data collection also requires substantial recruitment, sustained participation, technical infrastructure, and context-specific labeling.Source: §2.3 · pp. 5–6

2.4.1 Cultural Diversity as Social Practices.

  • The paper frames culture through social practices, distinguishing individual practitioners from community-level practices and their recurring, socially recognized patterns.Practices are described as involving material resources, competence, and meaning.Source: §2.4.1 · pp. 6–7

2.4.2 Capturing Social Practices.

  • DiversityOne combines time-use diaries and experience-sampling perspectives with smartphone sensing, allowing researchers to relate sensor patterns to personal and social context.Self-reports can provide ground-truth events, while sensor information can also help validate user annotations.Source: §2.4.2 · p. 7

2.5 Currently Available Public Smartphone Sensing Datasets

  • Existing public datasets vary in duration, scale, geography, self-reporting, and sensor coverage; GLOBEM is longitudinal but predominantly US-focused, while LifeSnaps covers four European countries with 71 participants.DiversityOne is described as spanning eight countries with 782 participants and offering finer-grained raw sensor data and frequent in-situ reports.Source: §2.5 · pp. 8–9

2.6 The DiversityOne Dataset

  • DiversityOne combines standardized behavioral and psychosocial measures with smartphone sensors, and emphasizes social practices, international breadth, and underrepresented Global South regions.The authors position it as a resource for cross-cultural and multidisciplinary analysis, not solely as a machine-learning benchmark.Source: §2.6 · pp. 9–10

3 Methodology

DiversityOne’s methodology pairs cross-site comparability with local adaptation, using questionnaires, intensive longitudinal diaries, and passive smartphone sensing to capture social practices and everyday behavior over time.

3 Methodology

  • The protocol was applied consistently across sites, with minor adjustments designed to preserve comparability while addressing local perspectives and requirements.Materials were adapted by local partners and validated by the initial team or an authorized body.Source: §3 · pp. 10–11

3.1 Questionnaires about Demographics and Social Practices

  • Questionnaires measured social practices through their materials, competencies, and meanings, covering relationships, activities, commuting, shopping, and cooking.The questionnaire was divided into three parts and administered during data collection.Source: §3.1 · pp. 11–12

3.1.1 Questionnaire Translation, Testing, and Adaptation.

  • Questionnaires were translated and locally adapted, then tested with approximately 30 participants to validate translations and assess completion times.Standard scales used validated translations where available; otherwise, expert translation and panel and back-translation were used.Source: §3.1.1 · pp. 12–13

3.2 Intensive Longitudinal Surveys and Passive Sensor Data Collection

  • An adapted iLog app collected intensive longitudinal self-reports alongside passive smartphone sensor data.The study design combined questionnaires with an intensive longitudinal survey approach to measure behavior over time.Source: §3 · p. 11, §3.2 · p. 13

3.2.1 Diary Methods and Daily Routines.

  • Daily diaries included morning and evening reports, time diaries, and snack diaries to record routines and activities.Time diaries asked about activity, location, social company, and mood; prompts ran every 30 minutes for the first two weeks and hourly for the final two weeks.Source: §3.2.1 · pp. 13–14

3.2.2 Sensor Data.

  • The sensor collection covered continuous and interaction modalities, including connectivity, environmental, motion, position, app-usage, and device-usage data.Environmental hardware sensors were available only on compatible devices.Source: §3.2.2 · pp. 14, 16

3.2.3 Intensive Longitudinal Survey Translation, Testing, and Adaptation.

  • Local testing and adaptation of iLog included a two-week app trial by approximately 15 participants and additional food questions tailored to partner interests.An offline version was developed for JLU and also used at AMRITA and IPICYT, where Google services or reliable data connections were unavailable.Source: §3.2.3 · p. 16

3.3 Target Population and Sampling Strategy

The study recruited college students across eight sites, adapting recruitment, incentives, and data-handling procedures to local circumstances while maintaining a shared collection protocol. Data ownership remained with local controllers, and anonymization reduced—but did not eliminate—re-identification risk.

3.3 Target Population and Sampling Strategy

  • College students were selected as the target population because shared student routines support comparisons of social practices across study sites.Sampling sought balance by gender, age, and academic department; students with incompatible phones or low survey attendance were excluded.Source: §3.3 · p. 16

3.3.1 Recruitment Strategy. / 3.3.2 Recruitment Adaptations.

  • Recruitment used invitations and reminders, with snowball recruitment at JLU and IPICYT and local adaptations to channels and timing.The study selected 350 students to target at least 250 active participants, accounting for expected longitudinal-study dropout.Source: §3.3.1 · p. 16, §3.3.2 · p. 17

3.4 Incentive Strategy

  • The incentive strategy emphasized fixed payments for completing at least 85% of notifications, supplemented by daily and final prizes.Incentives were adapted to local practices and cost of goods; some sites instead used certificates or social engagement.Source: §3.4 · p. 17, §3.4.1 · p. 17

3.4.1 Incentives Adaptation.

  • AMRITA’s certificates and IPICYT’s social-engagement approach worked for questionnaires but not as well for the intensive longitudinal survey.Incentives were adjusted to local costs and circumstances; at AMRITA, the approach’s novelty may also have contributed to the weaker survey outcome.Source: §3.4.1 · p. 17

3.5 Ethics and Privacy

  • Each country appointed a local data controller and ethics committee, while European data processing followed a GDPR-based framework adapted to local requirements.For IPICYT, participant requests for data deletion had to be fulfilled within three days.Source: §3.5 · p. 18

3.6 Data Collection

  • Collection took place during academic semesters and followed a shared sequence of invitation, selection, two collection phases, and closing procedures.Collection began several months later than planned in AMRITA and IPICYT, mainly because of technical challenges and ethics approvals.Source: §3.6 · pp. 18–19

3.7 Data Management / 3.7.1 Data Ownership and Copyright.

  • Local data controllers retained ownership, while UNITN processed the data and shared cleaned copies with controllers and researchers involved in refinement and analysis.Controllers later licensed their cleaned and anonymized datasets to UNITN under shared management and distribution policies.Source: §3.7.1 · p. 19

3.7.2 Data Cleaning and Anonymization.

  • Anonymization removed personal identifiers, anonymized network data, and truncated high-precision GPS data, but cross-referencing could still enable re-identification.The dataset is therefore available only under specific conditions rather than publicly online.Source: §3.7.2 · p. 19

4 Validation

Validation describes study engagement and data provision, then illustrates how everyday activities, food, mood, and app use vary across pilot sites.

4.1 Overall Study Engagement

  • 18,628 students responded to the initial questionnaire, and 782 installed the iLog app for data collection.Participation was highest at NUM and UNITN and lowest at AAU and IPICYT; up to 350 students per site were randomly selected with gender and department considered.Source: §4.1 · pp. 19–20

4.2 Response Rate in Time Diaries

  • In all pilots except AMRITA, 40% of participants averaged 20 daily responses, roughly ten hours of annotated sensor data per day.IPICYT’s higher response rate was partly attributed to a study configuration with more questions; participation consistency mainly varied with dropout.Source: §4.2 · pp. 20–21, Figure 3 · p. 21

4.3 Sensor Data Provision

  • AMRITA sensor provision fell from 80% on day one to under 25% by the final day, while average dropout across sites was approximately 35%.JLU and IPICYT initially rose because some participants joined after data collection began.Source: §4.3 · p. 21, Figure 4 · p. 21
  • Sensor participation varied by modality, and a lower participant percentage did not necessarily indicate lower data quality.Fixed-interval sensors such as accelerometers and application usage tended to have more participants with data than on-change sensors or pressure.Source: §4.3 · p. 21, Figure 5 · p. 22

4.4 Diversity in Everyday Life

  • Daily activity patterns differed across sites: many JLU students slept in the afternoon, and study annotations at 8 PM exceeded 30% in JLU versus 7% in IPICYT.Sleep annotations were primarily collected through an app feature for defining sleep windows, followed by regular morning self-reporting.Source: §4.4 · p. 23, Figure 6 · p. 23
  • Eating times varied by site, while water ranked first among drinks at UNITN but 16th at NUM and JLU, and appeared among top-reported snack-break drinks across all sites.AMRITA had eating spikes around 8 AM and 10 PM, NUM around 11 AM, and UC around 6 PM.Source: §4.4 · pp. 23–24, Figure 7 · p. 23
  • Mood reports showed similar daily trends but different absolute percentages across sites, with UC lowest and AMRITA highest; ratings also differed between home and university settings.The reported measure is valence, and personal rating criteria influence between-site differences; social networking and messaging apps were common across universities, with JLU students favoring WeChat.Source: §4.4 · p. 24, Figure 8 · p. 24, Table 5 · p. 24

5 Dataset and Code Availability

DiversityOne data are distributed through a secure, GDPR-compliant catalog, with selectable datasets and bundles, documentation, and a research-request process.

5 Dataset and Code Availability

  • Because the data include varied and sensitive sources, access is provided through a secure environment compliant with GDPR.The catalog is the main entry point for dataset documentation and requests.Source: §5 · p. 25

5.1 Downloadable Datasets and Bundles

  • Researchers can request single-site datasets, multi-site combinations, thematic bundles, or the full dataset, which is approximately 94 GB.Thematic bundles group commonly used data, such as motion sensors for activity recognition or questionnaires, time diaries, and location data for social-interaction studies; datasets are provided in Parquet format.Source: §5.1 · p. 25

5.2 Metadata and Documentation

  • Each dataset and bundle includes metadata and documentation covering collection, content, codebooks, related projects, and request procedures.The catalog supports searching by metadata and provides information on variables and labels to support analysis and replication of collection procedures.Source: §5.2 · p. 26

5.3 Dataset Request Process

  • Institution-affiliated researchers request data using selected catalog identifiers and a research proposal, then sign a license agreement if approved.The agreement restricts use to research, prohibits redistribution or public sharing, and forbids attempts to re-identify participants.Source: §5.3 · p. 26

6 Discussion

The discussion draws lessons from cross-country data collection and prior DiversityOne analyses, emphasizing adaptable protocols and models that account for local, individual, and multimodal variation. It also identifies participation and communication challenges that qualify the dataset’s collection.

6.1 Data Collection – Lessons Learned and Recommendations for the Future

  • Cross-country collection benefits from a GDPR-compliant core protocol adapted to local requirements, offline options for limited connectivity, and centralized documentation.The authors recommend adaptable protocols spanning study stages, from protocol design to anonymization, to support responsible future studies.Source: §6.1 · pp. 26–27

6.2 Dataset analysis – Lessons Learned and Recommendations for the Future

  • DiversityOne analyses highlight how local context, country-specific nuances, and multimodal data streams affect mobile-sensing model effectiveness.The section presents lessons from prior studies alongside recommendations for future research.Source: §6.2 · p. 28

6.2.1 Country-specific Models over Multi-country Models.

  • Country-specific models outperform multi-country models in reported tasks, while region-specific sensor patterns challenge cross-country generalization.The discussion contrasts this with computer-vision models’ effective generalization across multi-country visual datasets and notes that behavioral time-series models struggle with regional nuances.Source: §6.2.1 · p. 28

6.2.2 Value of Hybrid Model Personalization.

  • Hybrid personalization improves behavioral inference by combining population patterns with individual features, while fully personalized models require more user-specific data.Mood models using individual app-usage and movement data capture variations that generic models may miss.Source: §6.2.2 · p. 29

6.2.3 Advancing Techniques in Multimodal Integration for Mobile Sensing.

  • Combining activity, location, Bluetooth, app usage, device use, and WiFi can capture broader contextual patterns, but advanced multimodal methods remain underused.The discussion cites M3BAT’s gains on specific classification tasks while calling for models that account for regional, cultural, and individual behaviors.Source: §6.2.3 · p. 29

6.2.4 The Challenge of Distribution Shift.

  • Country-level sensor distributions can shift substantially; similar GPS patterns may indicate social activity in one country and commuting in another.The discussion reports Cohen’s d values up to 1.0 for sensing modalities including location and step count, and notes that multimodal domain adaptation remains nascent.Source: §6.2.4 · pp. 29–30

6.2.5 Effect of Label Quality on Model Performance.

  • Subjective self-reports such as mood show lower consistency and reliability than less subjective social-context labels, motivating more reliable labeling approaches.Suggested approaches include regularizing outliers, weighting conflicting reports, and using attention mechanisms to emphasize higher-confidence data.Source: §6.2.5 · pp. 30–31

6.3 Limitations

  • Collection faced communication challenges, uneven recruitment responses and dropout, respondent burden, and incentive disparities across sites.The authors connect these challenges to cultural differences, support delays, and extensive study materials, and recommend adapting recruitment and engagement practices.Source: §6.3 · pp. 31–32

7 Conclusion

DiversityOne addresses gaps in cross-country smartphone sensing with a publicly released, multidisciplinary dataset of student behavior and self-reports. Its scale, raw sensor data, and study documentation are intended to support diverse analyses and future diversity- and privacy-aware research.

7 Conclusion

  • DiversityOne combines passive smartphone sensing and self-reports from 782 college students across eight countries over four weeks, with 26 sensors and more than 350,000 reports.The paper describes it as one of the largest and most diverse publicly available datasets of its kind.Source: §7 · p. 32
  • Its interdisciplinary design draws on computer science, behavioral sciences, social practice theory, and intensive longitudinal time diaries to support inference about everyday behavior.The paper names mood, activities, social context, and eating habits as examples.Source: §7 · p. 32
  • The release includes raw sensor data and documentation of study design, collection, lessons learned, and future directions for diversity- and privacy-aware studies.The paper presents the dataset as a resource for ubiquitous computing, human-computer interaction, and machine learning.Source: §7 · p. 32

8 Author contributions

The authors list contributions by institution and then by individuals within each institution, and identify roles spanning study management, design, support, collection, preparation, analysis, and writing.

8 Author contributions

  • Author-name order follows contribution by institution and then by individuals within each institution, so it does not necessarily indicate individual contribution importance.Roles are identified by author initials.Source: §8 · p. 32
  • Study management is credited to F.G., I.B., A.D.G., and M.B.; study design to F.G., I.B., G.G., A.D.G., M.B., R.C.A., L.M., and D.G.P.Source: §8 · pp. 32–33
  • The listed roles also include technical support, data collection, data preparation and correction, data analysis, and manuscript writing.A.B. and L.M. are credited with data analysis; the manuscript-writing list includes M.B., A.B., L.M., D.G.P., and F.G.Source: §8 · p. 33

A.1 Time Diaries

The time diaries combine scheduled prompts with morning and evening questionnaires, collecting reports on daily activities and related context at varying intervals across the study.

A.1 Time Diaries

  • Morning diaries sent at 8:00 AM asked about the previous night’s sleep quality and expectations for the upcoming day.Evening diaries sent at 10:00 PM asked participants to evaluate the day, report problems and responses, and describe pandemic-related challenges.Source: §A.1 · p. 39, Table 10 · p. 40, Table 11 · p. 40
  • Participants could pause fixed-interval questions for a selected period to reduce answering burden.Source: §A.1 · p. 39
  • Prompts asked about activity, location, social context, and mood every 30 minutes for the first two weeks and hourly for the final two weeks.Conditional questions asked about travel purpose and transport mode, food and drink types, or sport type, depending on the reported activity.Source: §A.1 · p. 39
  • Separate prompts collected information about snacks and drinks outside main meal periods every two hours.Source: §A.1 · p. 39

A7. How was your day? A8. Did you have any problem at [college (weekdays)] today?

The questionnaire asks whether the participant could solve a problem, alone or with someone else.

A7. How was your day? A8. Did you have any problem at [college (weekdays)] today?

A9. What was the problem you had?

This appendix item asks what problem the participant had and whether COVID-19 prevented an activity they wanted to do.

A9. What was the problem you had?

  • Participants were asked whether COVID-19 made an activity they wanted to do that day impossible.The item is phrased as an open-ended question.Source: A9. What was the problem you had? · p. 40
  • A separate table lists motivations for suspending question notifications for a fixed number of hours.Source: Table 12 · p. 40

A.2 Sensors Description

The sensor appendix describes iLog’s smartphone sensor categories, collected variables, and recording cadence, covering both hardware measurements and software-generated events.

A.2 Sensors Description

  • iLog sensors are grouped as hardware sensors that detect physical conditions and software sensors that collect operating-system or software events.The sensor categories correspond to the distribution bundles.Source: §A.2 · pp. 40–41, Table 15 · p. 43
  • Collection cadence is described as on-change, up to a maximum number of samples per second, or once every specified interval.Actual sampling can diverge slightly from the reported frequency because of device inaccuracy.Source: §A.2 · p. 41
  • Table 16 describes the collected iLog sensors and their variables.Source: Table 16 · p. 44
  • Software events cover connectivity, device state, screen interaction, notifications, application use, and environmental readings such as light and pressure.Examples include detected Bluetooth and Wi-Fi networks, charging status, screen touches, and notification events.Source: §A.2 · pp. 44, 46
  • Hardware measurements include motion and position signals such as acceleration, rotation, activity, steps, location, and magnetic field.The accelerometer and gyroscope report 3D vectors; location includes coordinates and related fields.Source: §A.2 · p. 45

A.3 Dataset Files Structure

Dataset files are arranged by collection site and data type, with a matching file linking participant identifiers between sensor records and surveys.

A.3 Dataset Files Structure

  • The folder structure organizes data by site, with separate areas for interactions, surveys, and sensor categories such as app usage, connectivity, motion, and position.The listed files use Parquet and CSV formats.Source: §A.3 · pp. 48–49
  • The file matching.csv maps participant identifiers in sensor data to the identifiers used in surveys.Source: §A.3 · p. 49

Related papers

Loading 2502.03347v1…