Affective Contagion in Social Systems

• Affective contagion is the process by which one individual's emotional state influences another, measurable through metrics like Granger-causality and synchrony.
• Research integrates dyadic, group, and network models to reveal mechanisms such as super-spreader effects, phase transitions, and differential tie-strength influences.
• Empirical approaches utilize multimodal annotation and computational simulations to predict and modulate emotion spread in contexts from workplace dynamics to digital platforms.

Affective contagion refers to the transference, induction, or synchronization of emotional states between individuals or across groups, mediated by expressive cues, social network structure, and interaction dynamics. In contrast to mere detection or classification of affect, contagion captures how one agent’s expressed or latent emotional state evokes correlates, matches, or influences another’s, resulting in observable temporal dependencies and often collective mood alignment. This phenomenon is foundational to empathy modeling, crowd psychology, workplace morale, social media dynamics, collaborative computing, and increasingly, artificial agent interaction.

1. Formal Definitions, Models, and Metrics

The foundational definition of affective contagion is the process by which “one individual’s affective response influences the felt emotions of another,” operationalized via explicit temporal alignment, directionality, and dynamic dependency between emotional states. In quantitative terms, contagion is typically measured through:

• Directional Influence: Granger-causal tests are standard, where for time series $v_A(t)$ and $v_B(t)$, one fits an autoregressive model for $v_B$ with and without $v_A$’s lagged history. The improvement in predictive power quantifies the influence (see WELD (Sun, 17 Oct 2025), The Neuroticism Paradox (Sun, 16 Oct 2025)).
• Synchrony and Convergence: The increase in temporal similarity (cross-correlation, mutual information, or dyadic synchrony features) between individuals is used as a proxy (see “Dynamics of Collective Group Affect” (Prabhu et al., 2024)).
• Diffusion Models: SIS-type and agent-based models simulate emotional transmission over networks, parameterized by contagion probability, tie-strength preference, and recovery mechanisms (see (Fan et al., 2020, Fan et al., 2016, Fan et al., 2017, Ma et al., 2024)).
• Group Annotation: Group-level affect at time t is estimated via weighted aggregators or consensus annotation in time windows (see (Prabhu et al., 2024)).

A summary of primary metrics:

Metric	Definition	Application Area
CCC	Concordance Correlation Coefficient ($v(t),\hat v(t)$)	Dyadic affect (OMG-Empathy (Barros et al., 2019))
Granger F	Log ratio of residuals, edge weight $F_{i\to j}$	Workplace, organization (Sun, 17 Oct 2025)
Synchrony ($\rho$)	Max cross-correlation in signals	Group convergence/divergence (Prabhu et al., 2024)
Diffusion velocity	Rate of infection spread (SI/SIS)	Social network cascades (Fan et al., 2016)
Cascade coefficient	Correlation of initial vs final polarity in thread	Asynchronous digital threads (Kraishan, 4 Nov 2025)

These models are further expanded by incorporating agent heterogeneity, higher-order group interactions, and trait/role-based moderation (e.g., “firewalls,” initiators, absorbers in LLM networks (Xu et al., 13 Dec 2025), signed simplicial complexes (Ma et al., 2024)).

2. Dyadic, Group, and Network Mechanisms

Affective contagion manifests through layered mechanisms contingent on interaction topology, signal modality, and agent properties.

• Dyadic (one-on-one): In controlled experimental paradigms, as in the OMG-Empathy dataset (Barros et al., 2019), a speaker’s narrative induces time-varying shifts in listener valence. The primary measure is how well models driven by speaker (and listener) behavior can predict the listener’s continuous affective ratings. Personalized vs. generalized empathy protocols dissect individual vs. shared trajectory recovery.
• Small Group: In workplaces (WELD (Sun, 17 Oct 2025), The Neuroticism Paradox (Sun, 16 Oct 2025)), directed affective influence networks emerge, with “super-spreaders” (typically high-neuroticism, low-conscientiousness nodes) exerting disproportionately large reproduction numbers ($R_0 \sim 16$), and high clustering coefficients ($C\sim0.7$) providing containment (“emotional quarantine zones”).
• Large-Scale Social Media: On platforms like Twitter, Weibo, or GitHub, contagion is structured by exposure (feed composition) and network ties. Positive emotions propagate predominantly within strong-tie clusters, while negative emotions—especially anger—exploit weak links, traversing disparate communities and achieving greater velocity and reach (Fan et al., 2016, Fan et al., 2020, Fan et al., 2017, Kraishan, 4 Nov 2025). Susceptibility is stratified, with “highly susceptible” users over four times more likely to adopt positive than negative affect (Ferrara et al., 2015).
• Group Interactions and Higher-Order Effects: The signed simplicial contagion model (Ma et al., 2024) demonstrates how group emotional interactions, especially within balanced triangles, can induce discontinuous phase transitions, bistability, and hysteresis. The proportion of negative edges (distrust, antagonism) modulates the transition from abrupt outbreaks to gradual, SIS-like diffusion.

3. Experimental and Computational Approaches

Empirical investigation leverages multi-modal data collection, annotation, and simulation:

• Multimodal Annotation: Audio-visual recording, facial expression analysis, joystick-based continuous rating (Barros et al., 2019), EEG hyperscanning with contrastive graph learning (fGCL + DGC (Huang et al., 2023)), and skin conductance sensors for frisson events (Huang et al., 2024) form the backbone of experimental designs.
• Temporal Protocols: Fine-grained windowing (e.g., 15 s (Prabhu et al., 2024)) enables measurement of momentary affect fluctuations and group-level synchronization/convergence. EWE (Evaluator Weighted Estimator) aggregates annotator judgments for ground-truth group affect.
• Sentiment Algorithms: LIWC, SentiStrength, deep multimodal models, and pre-trained LLMs label emotional loading, sentiment, and polarity.
• Simulations and Network Analysis: Agent-based models, SI/SIS diffusion, and VAR/Granger-causality network reconstruction allow quantification and visualization of cascade processes. Signed networks, backbone extraction, and centrality metrics support structural analysis (Ma et al., 2024, Fan et al., 2017, Sun, 17 Oct 2025).

4. Moderators: Personality, Structure, Modality

Contagion susceptibility and impact are sharply moderated by:

• Personality Traits: High neuroticism yields larger out-degree in Granger-causal influence networks, while extraversion and conscientiousness are less predictive (Sun, 16 Oct 2025). Trait affectivity reverses directionality—high trait individuals may be pushed away from certain affective states, while low-trait individuals are attracted (Alshamsi et al., 2015).
• Social Structure: Weak ties enable rapid cross-cluster transmission of anger; strong ties limit spread to local communities (Fan et al., 2020, Fan et al., 2016). Group interactions induce nonlinearity and phase transitions when internal coherence is high, but collapse in presence of distrust (Ma et al., 2024).
• Modality: Face-to-face contagion exploits nonverbal signals (facial, prosodic); text-only contagion relies on semantic and emotional tagging (Ferrara et al., 2015, Kraishan, 4 Nov 2025). Machine agents (LLMs) can both instantiate and transmit structured affective dynamics (“chain-of-affective”)—role specialization further modulates group outcomes (Xu et al., 13 Dec 2025). Linguistic signals suffice for propagation in online environments, with explicit attention to emoji channels amplifying positivity (Kraishan, 4 Nov 2025).

5. Applications: Collaborative Systems, Organizational Health, Digital Platforms

• Affective Computing and Artificial Empathy: Systems are moving from individual-level reaction to coordinated modulation (group-level orchestration, policy learning) (Koch et al., 16 Jul 2025). Generative agents can sense group mood and inject responses to steer trajectories. LLMs exhibit measurable chain-of-affective processes that can be tuned for desired interaction outcomes (Xu et al., 13 Dec 2025).
• Workplace Dynamics: COVID-era datasets enable tracking emotional contagion through crises, mapping hubs of influence, and linking emotional climate to turnover prediction and event modulation (Sun, 17 Oct 2025, Sun, 16 Oct 2025).
• Crowd Management: Kinetic-theory models demonstrate how fear contagion accelerates evacuation, with behavioral trade-offs between individualistic (search) and herding (stream-following) strategies (Kim et al., 2020).
• Online Video and Code Collaboration: Physiological feedback sharing (frisson detection) and emoji reactions create new low-effort propagation channels, shaping collective enjoyment, social presence, and collaborative outcomes (Huang et al., 2024, Kraishan, 4 Nov 2025).

6. Limitations, Extensions, and Future Directions

Key limitations and open problems include:

• Causal Identification: Most models are observational; Granger-causality cannot isolate contagion from shared context or latent confounders (Sun, 17 Oct 2025, Ferrara et al., 2015). Controlled randomization is ethically fraught.
• Trait × State × Structure Interactions: Simple contagion models often disregard stable dispositional moderators and complementarity/adaptation processes (Alshamsi et al., 2015).
• Multi-Modality and Annotation: Single-modal proxies (facial, text) may miss multi-channel dynamics; advancing multimodal annotation and synchrony metrics is ongoing (Prabhu et al., 2024, Huang et al., 2023).
• Higher-Order and Long-Term Effects: Most studies focus on immediate emotional transfer; the persistence, cumulative impacts, or regulatory interventions remain understudied.
• Algorithmic and Ethical Risks: Steering group emotions via agent orchestration invokes concerns of manipulation and cultural bias—demanding transparency, audit, and user autonomy (Koch et al., 16 Jul 2025, Xu et al., 13 Dec 2025).

Future research must address:

• Causality-aware controllers and intervention modeling (instrumental variables, difference-in-differences).
• Real-time contagion detection and early warning in high-volatility scenarios (cyberbullying, outrage).
• Privacy-preserving, fairness-checked deployment of affective systems in real-world organizations.

7. Theoretical Synthesis and Implications

Affective contagion is best conceptualized as a dynamical process operating at multiple scales and modulated by agent traits, network topology, and interaction context. Classic pathogen-based models (SIS/SIR) provide a minimal formal substrate, but richer frameworks—Affective Epidemiology, signed simplicial contagion, chain-of-affective in LLM ensembles—capture the nonlinear, bistable, trait-dependent and role-specialized reality of human and artificial emotional propagation.

By integrating direct measurement, network modeling, and agent-based simulation, current research supports precise estimation, prediction, and eventual modulation of collective affect, with substantial implications for collaborative intelligence, organizational health, public policy, and AI system alignment.