TrustmeWatcher: An Application for Workplace Micro-Sensing and Explainable Well-Being Feedback

1 Introduction

• The project aims to study how sensing and AI can support knowledge workers while preserving privacy, autonomy, and understandable explanations.TRUST-ME brings together Università della Svizzera italiana and the Jožef Stefan Institute.Source: §1 · p. 1
• Computer activity and self-reports capture different perspectives, and neither should stand in for the other.Activity logs describe computer use, while self-reports describe perceived states and productivity.Source: §1 · p. 1
• TrustmeWatcher reuses ActivityWatch watchers and adds a service and interface for activity review, questionnaire synchronization, labeled records, and local prediction.The first study phase collected activity and repeated self-reports from 17 participants; the second phase will use the participant-facing application.Source: §1 · p. 1
• The application combines an interactive screen-time dashboard, self-report collection, privacy controls, daily predictions, and a counterfactual interface for repeated explanation encounters.The longitudinal field-study record connects repeated self-reports from 17 participants with surrounding computer activity.Source: §1 · p. 2

2 Related Work

• StudentLife paired continuous phone sensing with ecological momentary assessment but withheld feedback, while RADAR-base supports sensing, questionnaires, study management, storage, and visualization.TrustmeWatcher instead lets participants review recorded activity, control ambient sensing, and receive estimates and explanations in the same application.Source: §2 · p. 2
• Desktop telemetry and daily productivity ratings provide distinct evidence, and TrustmeWatcher aligns self-reports with surrounding computer activity.Beller et al. found that activity measures cannot replace a person’s account of the day [2].Source: §2 · p. 2
• Counterfactual generation and accessible explanations have prior examples, but explanations must fit users’ contexts and questions.Repeated encounters raise questions about comprehension, trust calibration, and changing mental models.Source: §2 · p. 2

3.1 System Roles and Data Flow

• ActivityWatch watchers record computer activity on the user device, while TrustmeWatcher’s local backend unifies activity history and supports synchronization, privacy controls, and local inference.Model development is outside the runtime boundaries; the trained model is deployed to the user device.Source: §3.1 · p. 2

3.2 A Unified Workplace Study Workflow

• The workflow presents activity across hourly to custom periods, groups daily video highlights, and places camera and eye-tracker pause/resume controls beside returned feedback.The screen-time summary combines a timeline, application distribution, and ranked category and application lists.Source: §3.2 · pp. 2–3

3.3 From Study Signals to Model Estimates

• Seven regressors estimate six state scores and overall well-being from ActivityWatch features, with questionnaire responses supplying targets and two collected dimensions excluded as targets.Physical comfort and social experience are collected but not predicted; local inference presents seven estimates on a common 0–6 scale.Source: §3.3 · p. 3
• Targets combine questionnaire items into distinct states, and regressors use 25 features summarizing the preceding hour plus 24-hour and seven-day context.Features include activity-state shares, application-category ratios, focus fragmentation, input activity, and changes from recent history; standardization is participant-specific when training data are available.Source: §3.3 · p. 3

3.4 From Model Estimates to Explanations

• State cards use low, moderate, and high bands, and selecting a card opens a counterfactual search over five-minute activity swaps.The search returns the valid allocation whose model estimate is closest to the user-selected target, preserving total time across categories.Source: §3.4 · p. 3

4.1 Data-Collection Study

• The archive contains pseudonymized records from 17 participants, spanning July 2025 to March 2026, with 10–28 workdays per participant and a median of 18.Participants worked in everyday settings rather than a single laboratory session.Source: §4.1 · p. 4
• Participants completed StreamDeck check-ins during work and rated the same dimensions while reviewing available video highlights, producing immediate and retrospective self-reports.All 17 contributed computer-activity context and check-ins; nine also contributed video questionnaire responses, aligned with preceding activity context while preserving source and timestamp.Source: §4.1 · p. 4

4.2 Prediction Performance

• Chronologically ordered questionnaire days were split approximately 80/10/10 for training, validation, and test, with boundary days excluded to reduce temporal overlap.The model was selected by validation MSE and evaluated once on held-out data, against random, global-mean, and personal-mean baselines.Source: §4.2 · p. 4

• Reported evaluation covers prediction performance and data-collection coverage.Source: §5 · p. 4
• Explanation comprehension, trust calibration, and interpretation of model scenarios require analysis of repeated participant interactions.Source: §5 · p. 4
• Computer activity omits meetings and off-screen work, while application categories cannot establish intent.Estimates therefore reflect recorded context rather than complete accounts of work or well-being.Source: §5 · p. 4

• TrustmeWatcher integrates a model developed from records collected from 17 participants with tools for the second study phase.Source: §6 · p. 4
• The application combines computer-activity review, self-report, privacy control, local prediction, and counterfactual explanations.Source: §6 · p. 4
• The integrated application provides infrastructure for longitudinal evaluation of explanation use, trust, and privacy choices.Source: §6 · p. 4