WELD: The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset for Ubiquitous Affective Computing

Technical Summary: WELD – The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset

Problem Statement
The authors identify a critical gap in affective computing: the absence of datasets that simultaneously satisfy four conditions necessary for ubiquitous, real-world deployment: (i) long-duration coverage (months to years), (ii) fully naturalistic, in-the-wild settings, (iii) stable small-team social structures, and (iv) fully passive sensing protocols that survive institutional review. Existing datasets typically offer only one or two of these axes (e.g., long duration but lab settings, or naturalistic settings but short duration and unrelated subjects). This limitation prevents the study of long-term within-individual emotional dynamics and team-level relational dynamics (e.g., emotional contagion) in authentic workplace contexts.

Methodology
The authors introduce WELD (Workplace Emotion Longitudinal Dataset), collected from 49 consenting employees of a Chinese software development company over 30.1 months (November 2021 – May 2024).

• Data Acquisition: The pipeline utilizes existing ceiling-mounted IP cameras (1080p, 25 fps) in an open-plan office. Processing occurs on-premises using a Python pipeline involving RetinaFace for detection, an internal embedding model for identity matching, and a commercial 7-class Facial Expression Recognition (FER) model fine-tuned on Asian faces. Sampling is passive and limited to work hours (08:00–20:00), with faces processed at most once every 10 seconds.
• Ethics and Privacy: The study received an IRB-exempt determination. A four-level desensitization scheme is employed:
  • L0/L1: Raw video and face crops/embeddings remain on-premises and are never released.
  • L2: Per-frame probabilities tied to internal pseudonyms are retained for 5 years but released only under strict Data Use Agreements (DUA).
  • L3: The public release consists of per-hour aggregated probabilities tied to randomized pseudonyms (P01–P49). It contains no images, demographics, or direct identifiers. A k-anonymity audit confirms a minimum group size of 8, and linkage attack simulations show success rates indistinguishable from chance.
• Cohorts: Of the 49 participants, 22 met the primary inclusion criterion of ≥60 active days (cohort_strict) for downstream analyses. The remaining 27 short-tenured individuals are retained for onboarding studies.
• Validation: System outputs were validated against two FACS-trained raters on a 1,000-frame subset, achieving 88.2% agreement (Cohen’s κ=0.84).

Key Contributions

1. Dataset Release: WELD is the first dataset to combine long-period, naturalistic, small-team, and passive sensing. It contains 733,780 per-frame records (725,676 in the cleaned L3 release).
2. Privacy Framework: A defensible, four-tier access model that balances reproducibility with participant protection, including a public demo subset (T0), reviewer access (T1), bona-fide research access via a Data Access Committee (T2), and never-released raw layers (T3).
3. Baseline Analyses: The paper provides reference analyses for five tasks: emotion recognition replication, cycle recovery, event-driven shift detection, turnover prediction, and FER bias auditing.
4. FER Bias Documentation: The first large-scale evidence of an "angry-on-Asian-neutral-face" bias in commercial FER systems, where the mean "angry" probability (0.194) is 4× higher than Western benchmark priors (0.05).

Results

• Variance Decomposition: Using Type-I sequential ANOVA on daily valence, the authors attribute variance as follows: 19.3% to stable between-person differences, 29.8% to month seasonality, 7.6% to a V2/V3 measurement pipeline transition, 3.3% to day-of-week, and 40.1% to within-person residuals. This establishes a performance ceiling for future models.
• Affective Phenomena Replication:
  • Weekend Effect: A +43.1% relative valence boost on weekends compared to weekdays.
  • Diurnal Cycle: A midday peak at 12:00 and a trough at 13:00 (amplitude ≈0.09 valence units), observed in 19 of 22 individuals.
  • Shanghai 2022 Lockdown: A naive comparison showed a marginally significant valence drop (d=−0.40, p=0.054). However, rigorous causal inference methods (interrupted time-series, permutation tests, SARIMA counterfactuals) found no significant negative causal effect, suggesting the naive result was confounded by a pre-existing downward trend.
• Turnover Prediction: Using only the first 30 days of emotion features, leave-one-person-out binary classification achieved an AUC of 0.79 (Gradient Boosting). However, when framed as a survival problem (Cox proportional-hazards), the concordance index dropped to 0.52 (chance level). The authors attribute the high AUC to "tenure-induced leakage" (short-tenured individuals are over-represented in the departed class) and recommend reporting both metrics to avoid methodological inflation.
• Emotional Regimes: A Hidden Markov Model (HMM) identified six emotional regimes. Negative regimes (e.g., "Hostile," "Subdued") exhibited asymmetric persistence, with dwell times 5–6× longer than the positive "Joyful" regime. Direct transitions from the most negative to the most positive state were essentially impossible.
• Contagion Networks: Granger causality analysis revealed a sparse network (2.7% density, 8 edges), while transfer-entropy analysis found a disjoint sparse network (8.1% density, 24 edges). Both methods indicate that dense contagion networks are not supported by the data.
• Periodicity: Lomb-Scargle periodograms confirmed multi-scale periodicity at 24 hours (circadian), 7.7 days (weekly), and 25–46 days (potentially project/pay-period cycles).

Significance
The paper claims WELD unlocks two scientific use cases previously inaccessible: long-term within-individual emotional dynamics and team-level relational dynamics. By providing a quantitative ceiling for variance explanation and demonstrating the confounding risks in naive event analysis (e.g., the lockdown study), WELD serves as a rigorous benchmark for future workplace affective computing. Additionally, by documenting the systematic over-prediction of "angry" on Asian neutral faces, the dataset provides a critical resource for fairness audits of facial emotion classifiers outside Western contexts. The authors emphasize that the conjunction of WELD’s four axes is difficult to replicate, requiring multi-year institutional commitment and strict ethical guardrails, and they advocate for this design as a template for future naturalistic workplace research.