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

arXiv:2510.15221 · cs.AI, cs.CY, cs.LG · Submitted 2025-10-17 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "WELD: The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset for Ubiquitous Affective Computing".

Jane: The paper was written by Xiao Sun from Hefei University of Technology and Anhui Province Key Laboratory of Affective Computing and Advanced Intelligent Machines, School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Summary: Tom: So, moving beyond the title, let's look at what "WELD: The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset for Ubiquitous Affective Computing" actually reveals about human feeling. The authors’ summary highlights some very clear patterns in how people feel in this workplace.

Jane: One key observation is the clear diurnal cycle; we see valence, or general feeling, consistently peaking around midday and dipping right after lunch every single day. This mirrors our own biological rhythms and shows us how our emotions are tied to the clock.

Lu: I am particularly fascinated by the evidence for the weekend effect; WELD shows a significant +forty-three point one percent relative boost in valence when comparing weekdays to weekends, which confirms that even in a structured environment, human happiness has these predictable seasonal rhythms.

Meng: From an operational standpoint, that means we could potentially build internal systems that recognize when it’s time for more focused work and less of the social interaction because the collective energy levels are changing. It helps us understand human capacity during work hours.

Lalam: Lalam thinks this highlights how vital it is to acknowledge these natural rhythms in our lives. If we can predict these shifts, we can design environments that support better well-being for everyone involved in that space.

Tom: These patterns are very real, but they also show some interesting moments of change too, especially during the Shanghai two thousand twenty-two lockdown period. The data showed a marginally significant drop in overall valence during that time.

Jane: That drop is quite subtle, but it’s important because it demonstrates how large-scale global events can affect even stable groups over months, something short-term studies would never be able to capture accurately.

Lu: It shows the power of the entire team; we are seeing how an external event impacts a single group consistently over time, not just a momentary snapshot of individuals.

Meng: It gives us data to ground our models in actual events, which is far better than relying on hypothetical scenarios when considering shifts like public health crises.

Lalam: We can use this knowledge to help design systems that are resilient or even adaptive when the world around them changes dramatically, supporting a more stable human experience during periods of uncertainty.

Improvements: Tom: The authors argue that "WELD: The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset for Ubiquitous Affective Computing" closes a massive gap in research by providing a whole new framework for understanding emotion. They aren't just offering data; they are providing a holistic approach to the field.

Jane: It’s about combining four elements—long duration, natural setting, small team structure, and passive sensing—that we usually find separately. This combination allows us to ask questions like "how does emotion spread through a stable team?" which is completely new territory for previous research.

Lu: The ability to study "within-individual" dynamics over years is a game changer; it lets us see if emotional volatility in someone's early career, for example, can predict future turnover. We can finally move beyond simple mood tracking and look at the history of an individual’s life.

Meng: And I appreciate the focus on the passive sensing protocol; it means that from a practical standpoint, we're not requiring users to wear special gear or fill out surveys. It just works by using existing office infrastructure for ubiquitous AI deployment.

Lalam: Lalam sees this as an improvement in our research methodology; we are learning how the human experience truly unfolds over years, allowing us to build systems that support a more holistic understanding of human development and interaction.

Tom: The authors also offer incredibly useful baselines for future modeling, particularly their work on turnover prediction. They found a significant gap between the binary AUC and the survival C-index which is extremely important.

Jane: That distinction is so crucial because it warns researchers against making overly optimistic claims based on simple predictive scores. It forces us to use more rigorous methods like survival analysis that actually track subjects over time.

Lu: The finding of six distinct emotional regimes, where negative states persist five times longer than positive ones, gives us a structural understanding of human emotion that we can feed into our models to make them much more sophisticated.

Meng: This allows us to build better algorithms because instead of treating emotions as random occurrences, we can model the inherent tendency for certain emotional cycles to drag on.

Lalam: We can use this structural insight to improve how we support workers, understanding that recovery from a negative state isn't instantaneous but requires a long process of transition and time.

Conclusion: Tom: So, we are wrapping up our discussion on "WELD: The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset for Ubiquitous Affective Computing," which has given us so much to talk about today. We have seen how this dataset provides a rigorous floor for future studies.

Jane: It’s clear that the authors are using this data to challenge old assumptions in Affective Computing, especially with the surprising findings regarding bias in facial expression models on Asian faces.

Lu: The ability to quantify things like emotional contagion and structural asymmetry is just so exciting; it allows us to build AI that understands not just what a person feels, but how they contribute to the collective dynamic of the team.

Meng: Knowing that we can model these patterns with a specific performance ceiling, according to the variance decomposition, means we have much clearer goals for building robust, real-world AI systems.

Lalam: Lalam hopes that this work paves the way for a more empathetic and culturally sensitive way to design technology, moving beyond the limitations of our current Western-centric models.

Tom: Before we go, I want to hear some final thoughts from our team members on what’s next or what this means moving forward.

Lu: We need more replication studies in different cultures because WELD is a single-culture study, and that’s a massive limitation we have to acknowledge.

Meng: I think the practical impact will be seeing how many companies are willing to adopt such rigorous, ethically sound data collection methods moving forward with their internal processes.

Lalam: Lalam believes the hope is that this data leads to better collaboration between different groups who are working toward a shared goal in a more supportive way.

Tom: That’s a great point, and I think we've covered everything from the science of emotion to the practical applications of this massive dataset. We're going to take a quick break, but when we come back, we'll be discussing another fascinating paper in Affective Computing.

Conclusion: Tom: So, we've been exploring WELD: The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset for Ubiquitous Affective Computing, and it’s clear this research has provided a massive toolkit for the field.

Jane: I think the biggest thing is that this dataset allows us to move beyond simple snapshots; we can finally see how emotion evolves over years in a way that's actually representative of real life.

Lu: The structural findings, especially about how negative emotional states persist much longer than positive ones, give us a deep understanding of human temperament that's essential for AI to grasp complex human behavior.

Meng: I am glad the system provides such clear benchmarks—it gives developers a concrete performance ceiling and an honest look at what AI can realistically achieve in a practical workplace setting.

Lalam: Lalam believes this work is helping us envision technology supporting more empathetic, culturally aware interactions, moving beyond the limitations of our current Western-centric models.

Tom: That's true; it really forces us to think about how the system itself is designed and what assumptions we are making about human nature.

Jane: And by highlighting issues like the "angry-on-Asian-neutral-face" bias, it is forcing a more rigorous look at fairness in AI applications that are being trained on limited data.

Lu: The fact that the team's collective dynamics can be seen through both linear and non-linear lenses means we are able to test the limits of how influence spreads within a bounded group.

Meng: Knowing that you can simulate those organizational shocks—like the two thousand twenty-two lockdown—allows us to build systems that handle unpredictable real-world events.

Lalam: I feel like this work is a powerful argument for creating better, more thoughtful consent frameworks for naturalistic research across all global contexts.

Tom: You're right, it’s a massive conversation starter about how we approach ethical data collection in the future.

Hefei University of Technology · Anhui Province Key Laboratory of Affective Computing and Advanced Intelligent Machines, School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China

cs.AI, cs.CY, cs.LG

Submitted: 2025-10-17

Updated: 2026-09-03

Comments: WELD: 733,780 per-frame 7-class facial-expression probability records from 49 employees over 30.1 months (Nov 2021-May 2024). v2 corrects attrition metrics after removing leakage (binary AUC=0.79, survival C-index=0.52) and adds a FER fairness audit. 49 pp, 14 figs, 1 supp PDF

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: As a diligent researcher whose work relies entirely on verifiable source material, I must report that the provided text consists solely of a reference list and page headers from an academic

Key concepts

Diurnal Cycle: WELD shows that general feeling (valence) consistently peaks around midday and dips after lunch every day. This pattern mirrors biological rhythms and demonstrates how emotions are tied to the clock in a naturalistic setting.
Weekend Effect: The dataset reveals a significant positive boost in valence when comparing weekdays to weekends. This suggests that human happiness has predictable, measurable seasonal rhythms even within structured environments.
Passive Sensing: This protocol allows for the ubiquitous deployment of AI by using existing office infrastructure to gather data. It avoids requiring users to wear special gear or complete surveys.

Terminology

Affective Computing — 35× in this paper · explained in 1 episode
AUC — 20× in this paper · explained in 2 episodes
Aggregation — 6× in this paper · explained in 1 episode
Embeddings — 6× in this paper · explained in 2 episodes
Benchmark — 5× in this paper · explained in 2 episodes
Calibration — 5× in this paper · explained in 7 episodes
Causal Inference — 3× in this paper · explained in 4 episodes
Accuracy — 2× in this paper · explained in 1 episode
Facial Action Coding System — 2× in this paper
Clustering — 1× in this paper · explained in 2 episodes
Commentary — 1× in this paper · explained in 2 episodes
Data Access Committee — 1× in this paper
Decode — 1× in this paper · explained in 2 episodes
Generalization — 1× in this paper · explained in 13 episodes
Representation Learning — 1× in this paper · explained in 2 episodes
Verification — 1× in this paper · explained in 2 episodes

Summary

As a diligent researcher whose work relies entirely on verifiable source material, I must report that the provided text consists solely of a reference list and page headers from an academic publication. To generate a summary of WELD: The First Naturalistic Long-Period Small-Team Workplace Emotion Dataset for Ubiquitous Affective Computing, I require the actual body content, abstract, or introduction sections of the paper itself.

The current input does not contain any descriptive text regarding the dataset's methodology, structure, findings, or significance that would allow me to construct a summary of 450–600 words while adhering strictly to quoting key phrases and avoiding external commentary. Please provide the main body text of the article for accurate summarization.

Improvements for AI systems

Improvement: Implement a federated, multi-stage inference pipeline that moves beyond simple correlation detection in affective data toward establishing temporally constrained, causal hypotheses about psychological state. This requires upgrading the current architecture to explicitly model the interaction between discrete life events and continuous biometric signals.

Technical Mechanism:

Graph-Based Temporal Modeling: Replace sequential RNN/LSTM processing with a Dynamic Knowledge Graph (DKG) structure. Nodes represent entities (individual, project, calendar event, detected emotion). Edges represent temporal relationships (precedes, concurrent with) and causal likelihoods (P(State B Event A)).
Multi-Modal Fusion Layer: Integrate a Transformer-based Fusion Attention Mechanism that weights inputs based on their reliability within the current context. For instance, if calendar metadata indicates Vacation, the attention weight for real-time stress signals (e.g., late-night activity) should be temporarily downweighted unless corroborated by other modalities (e.g., low sleep trackers).
Privacy-Preserving Edge Processing: Enforce a Differential Privacy (DP) constraint on all data leaving the local, bounded group system. Instead of transmitting raw affective vectors, the edge device calculates and transmits only gradients or aggregated statistical deviation profiles (mu) relative to the established baseline for that specific user/context pairing.

What the Improved AI System Can Do:

The CASM system can generate Actionable Affective Forecasts rather than mere real-time readings. Specifically, it can:

Identify Contextual Stressors (Causal Attribution): Determine not just that a user is stressed, but with high statistical confidence (p < 0.01), pinpoint the specific confluence of factors—e.g., The combination of the deadline for Project X (Metadata) occurring within 48 hours of insufficient sleep (Bio-signal) and conflicting priorities (Calendar/Task-tracker)—is the primary driver of elevated negative affect.
Proactive Interventional Suggestion: Based on longitudinal pattern recognition, the system can issue highly personalized, preemptive interventions. Instead of waiting for a mood dip, it predicts a high probability of negative affective drift 72 hours in advance and suggests specific mitigations (e.g., Warning: Low engagement predicted next Tuesday due to project overlap. Consider blocking 2 hours for focused work or rescheduling the meeting with Group B.).
Quantify Systemic Friction: Analyze the collective data from the bounded group to identify points of systemic friction within the organization that impact affective well-being (e.g., The current onboarding process consistently generates moderate anxiety peaks in new members 14 days post-start, suggesting a need for structural review).

Bias Auditing: It provides a quantitative Bias Score alongside every prediction. A low score indicates that the prediction relies on stable, universally recognized psychological constructs rather than superficial or culturally biased biometric markers.
Robustness Guarantee: The system can confidently report its Confidence Domain, stating not only what emotion it predicts, but also under which conditions (e.g., Predicted: Focused Concentration; Confidence Domain: High; Limitations: Requires sustained audio input and visible task-related metadata).
Explainability for Trust: It generates a detailed Attribution Map for every affective judgment, showing exactly which combination of evidence (e.g., 60% weight from Calendar Conflict, 30% from decreased typing speed, 10% from vocal pitch variance) led to the final conclusion, thereby building necessary trust and enabling human oversight.

Abstract

Affective computing has matured rapidly in laboratory settings, yet no prior dataset combines (i) months-to-years of duration, (ii) a naturalistic workplace context, (iii) a stable small-team social structure, and (iv) a fully passive sensing protocol that survives institutional review. We introduce WELD, the first dataset to satisfy all four. WELD comprises 733,780 per-frame seven-class facial-expression probability vectors from 49 employees of a Chinese software company over 30.1 months (Nov 2021 - May 2024) -- the longest naturalistic in-the-wild emotion corpus and the only multi-year corpus supporting both within-individual longitudinal and within-team relational analyses on the same subjects. Data are released under a four-tier access model with only aggregated probabilities publicly downloadable. We validate the corpus by replicating three established phenomena (+43.1% weekend valence boost; 13:00-trough diurnal cycle; Shanghai 2022 lockdown effect d=-0.40), and report four novel findings: (1) variance decomposition attributes 19.3% of daily-valence variance to between-person differences and 29.8% to month seasonality -- a quantitative ceiling for future predictive models; (2) Hidden Markov decomposition reveals six emotional regimes with asymmetric negative-state dwell times (16-18 d vs 3 d); (3) leave-one-person-out turnover prediction reaches AUC=0.79 yet a Cox concordance index of only 0.52, exposing a metric-trap when AUC is reported without survival-aware baselines; (4) the corpus reveals systematic over-prediction of "angry" by an off-the-shelf FER model on neutral Asian faces (0.194 vs 0.05 Western priors), making WELD valuable for FER fairness audits. A complex-systems analysis of the corpus appears as a companion preprint (arXiv:2510.16046).