Check out our working paper: A Survey of Social Network Simulation in the LLM Era: From Classical Models to Generative Agents View on SSRN
A collection of academic articles, published methodology, and datasets on the subject of Social Network Simulation from classical models to LLM-based agentic simulation.
- Taxonomy
- Network Models
- Information Diffusion
- Opinion Dynamics
- Rule-Based Agent-Based Models
- LLM-Based Agentic Simulation
- Datasets and Benchmarks
- Evaluation Metrics
- Contributing
If you use this repository, please cite the survey manuscript. A public DOI or finalized venue entry will be added once the paper metadata is finalized.
@article{nguyen6790227survey,
title={A Survey of Social Network Simulation in the LLM Era: From Classical Models to Generative Agents},
author={Nguyen, Thanh Tam and Pham, Trinh and Huynh, Viet and Nguyen, Minh Hieu and Vo, Bay and Qu, Liang and Li, Jianxin and Yin, Hongzhi and Nguyen, Quoc Viet Hung},
journal={Available at SSRN 6790227}
}
The survey organizes the field into five paradigm families:
- Network Models: topology generation and structural analysis for synthetic social graphs.
- Information Diffusion: epidemic, cascade, influence, rumor, and control models.
- Opinion Dynamics: averaging, bounded-confidence, polarization, and echo-chamber models.
- Rule-Based Agent-Based Models: heterogeneous bottom-up simulation with explicit agent rules.
- LLM-Based Agentic Simulation: persona-grounded, memory-enabled, language-driven social agents.
Each table below is grounded in the manuscript citations and keeps the paper title, year, category, venue, and repository or project link visible for quick scanning. Paper titles link to official publication or publisher pages when confidently resolved.
The survey catalogues 21 published datasets and benchmarks for validating structural fidelity, temporal dynamics, behavioral realism, and platform-scale effects. The table below highlights the main benchmark families surfaced in the paper appendix.
| Category | Benchmark / Dataset | Platform | Evaluation Task | Used In |
|---|---|---|---|---|
| Political Behavior and Elections | PPE Benchmark | Voter preference prediction | ElectionSim | |
| ANES 2020 | Survey | Demographic calibration | ElectionSim | |
| ITA-ELECTION-22 | Political conversation simulation | SOSMC | ||
| User Behavior and Engagement | Vaccination Engagement | X | Action-guided engagement | CLSSM |
| CCP-DE | X | Comment prediction (DE) | TSSMU | |
| CCP-EN | X | Comment prediction (EN) | TSSMU | |
| CCP-LU | RTL | Comment prediction (LU) | TSSMU | |
| Y Social traces | Synthetic | Digital-twin validation | Y Social | |
| Misinformation and Rumor | Rumor Event Suite | Temporal rumor dynamics | RumorSphere | |
| BotSim-24 | Adversarial bot simulation | BotSim | ||
| FPS Fake-News | Synthetic | Attitude dynamics to fake news | FPS | |
| FUSE Mutation | Synthetic | Stepwise news deception | FUSE | |
| Controversial-Topic Propagation | Gender Discrimination | Emotion and event propagation | S3 | |
| Nuclear Energy | Attitude-change propagation | S3 | ||
| Opinion Dynamics and Echo Chambers | LLM Opinion Dynamics | Synthetic | Opinion convergence | SODNL |
| OASIS Twitter-like | Synthetic | Group polarization | OASIS | |
| OASIS Reddit-like | Synthetic | Herd effect and misinformation | OASIS | |
| Network Structure | SNAP Collection | Multi | Network analysis and topology generation | NetworkX, igraph |
The survey groups 22 evaluation metrics into seven core evaluation goals. The table below summarizes the metric families that recur throughout the paper.
| Category | Metric | Formula / Description | Usage |
|---|---|---|---|
| Structural Fidelity | Degree distribution | P(k) = Pr(deg(v)=k); scale-free often fits P(k) ∝ k^{-γ}. |
Validates topology realism (heavy-tail, hubs) of generated networks. |
| Clustering coefficient | Local: C_i = 2T_i / (k_i (k_i-1)), where T_i is triangles incident to node i. |
Measures triadic closure; high clustering indicates community / small-world structure. | |
| Average path length | L = (2/(N(N-1))) · Σ_{i<j} d(i,j). |
Checks global reachability and small-world effect (L typically grows like log N). |
|
| Assortativity coefficient | Pearson correlation of endpoint attributes across edges; degree-assortativity often reported as r. |
Quantifies homophily/heterophily (e.g., r > 0 assortative, r < 0 disassortative). |
|
| Dynamic and Temporal Fidelity | Pearson correlation | r = cov(x, y) / (σ_x σ_y). |
Tests temporal synchronization between simulated and empirical trajectories. |
| DTW distance | DTW(X,Y) = min_π Σ_{(i,j)∈π} δ(x_i, y_j) over warping paths π. |
Sequence alignment with time-warping; used for cascade / activity curve matching. | |
| DeltaBias | ΔBias = Bias(sim) - Bias(emp) (or absolute gap). |
Measures directional error in simulated bias vs. an empirical baseline. | |
| DeltaDiversity | ΔDiversity = Div(sim) - Div(emp) (or absolute gap). |
Evaluates whether content / exposure diversity is preserved during propagation. | |
| Opinion and Polarization | Polarization P_z |
Variance proxy: P_z = (1/N) · Σ_i (x_i - x̄)^2. |
Tracks population-level opinion dispersion (convergence vs. fragmentation). |
| Global disagreement | Edge-averaged distance, e.g. GD = (1/|E|) · Σ_{(i,j)∈E} |x_i - x_j|. |
Captures cross-edge heterogeneity (echo-chamber separation yields higher disagreement). | |
| Neighbour correlation index (NCI) | NCI = (1/|E|) · Σ_{(i,j)∈E} 1[|x_i - x_j| < ε]. |
Quantifies local homogeneity (fraction of “similar” neighbors under tolerance ε). |
|
| Spectral gap | Often reported via the second eigenvalue, e.g. gap = 1 - |λ_2(W)| for mixing matrix W. |
Measures convergence/mixing speed; larger gap typically implies faster consensus. | |
| Diffusion and Influence | Reproduction number R_0 |
Basic ratio form (SIR-style): R_0 = β/γ (infection rate over recovery rate). |
Checks whether diffusion is supercritical (R_0 > 1) or dies out (R_0 < 1). |
| Influence spread | Expected activated count: σ(S) = E[|Activated(S)|] from seed set S. |
Evaluates seed effectiveness and reach under IC/LT-style cascade processes. | |
| Cascade size | Final activated / infected count (or its distribution) over runs. | Measures total reach of an information cascade and validates diffusion intensity. | |
| Behavioral Alignment | Micro-F1 | F1 = 2PR/(P+R) computed with micro-aggregated counts. |
Per-instance prediction quality (e.g., stance/action classification, voter prediction). |
| Action distribution divergence | KL: D_KL(P||Q)=Σ_a P(a) log(P(a)/Q(a)); JS is a symmetric alternative. |
Checks whether simulated engagement/action frequencies match real logs. | |
| Cosine similarity | cos(e_a,e_b) = (e_a·e_b)/(||e_a||·||e_b||) for embedding vectors. |
Semantic alignment between simulated content and reference content. | |
| State-level accuracy | Acc = (# correct states)/(# total states). |
Validates discrete outcomes (e.g., election results, topic/state transitions). | |
| Efficiency and Scalability | Confusion index | Heuristic for “hard” cases (model-dependent; typically combines uncertainty and disagreement signals). | Enables selective LLM invocation (route easy steps to cheap rules; hard steps to LLMs). |
| Token reduction | TR = Tokens(baseline) / Tokens(method) (or relative % reduction). |
Quantifies compute savings from hybrid scheduling, caching, summarization, etc. | |
| Wall-clock cost | Measured runtime / $ cost per simulation step or per episode. | Assesses feasibility at scale (million-agent runs, long horizons, repeated trials). | |
| Adversarial Robustness | Detector accuracy drop | ΔAcc = Acc(clean) - Acc(attack). |
Stress-tests defenses against LLM-driven bots and adversarial coordination strategies. |
Feel free to contact us if you have any queries or exciting news on social network simulation. In addition, we welcome all researchers to contribute to this repository and further contribute to the knowledge of social network simulation fields. It would be great if contributions keep the repository aligned with the survey taxonomy
If you have some other related references, please feel free to create a Github issue with the paper information. We will glady update the repos according to your suggestions. (You can also create pull requests, but it might take some time for us to do the merge)
Backup Statistics
