Constructing a Dataset to Support Agent-Based Modeling of Online Interactions: Users, Topics, and Interaction Networks

Abdul Sittar, Miha Češnovar, Alenka Guček, Jeff Z. PanarXiv 2026 · DOI · arXiv

Abstract

Agent-based modeling (ABM) provides a powerful framework for exploring how individual behaviors and interactions give rise to collective social dynamics. However, most ABMs rely on handcrafted or parameterized agent rules that are not empirically grounded, thereby limiting their realism and validation against observed data. To address this gap, we constructed a large-scale, empirically grounded dataset from Reddit to support the development and evaluation of agent-based social simulations. The dataset includes 33 technology-focused, 14 climate-focused, and 7 COVID-related aggregated agents, encompassing around one million posts and comments. Using publicly available posts and comments, we define agent categories based on content and interaction patterns, derive inter-agent relationships from temporal commenting behaviors, and build a directed, weighted network that reflects empirically observed user connections. The resulting dataset enables researchers to calibrate and benchmark agent behavior, network structure, and information diffusion processes against real social dynamics. Our quantitative analysis reveals clear topic-dependent differences in how users interact. Climate discussions show dense, highly connected networks with sustained engagement, COVID-related interactions are sparse and mostly one-directional, and technology discussions are organized around a small number of central hubs. Manual qualitative analysis further shows that agent interactions follow realistic patterns of timing, similarity between users, and sentiment change.

The Realism Gap in Social Simulation

Agent-based modeling (ABM) is a powerful tool for studying how individual behaviors lead to collective social phenomena like the spread of misinformation or the formation of echo chambers. However, most ABMs suffer from a "realism gap." They typically rely on handcrafted rules or simplified parameters that aren't grounded in actual human data, making it difficult to validate whether a simulation reflects reality or just the modeler's assumptions.

Furthermore, as the field moves toward using Large Language Models (LLMs) to power agents, there is a desperate need for datasets that capture not just what people say, but how they interact over time within specific social structures. This research addresses the lack of high-quality, simulation-ready datasets that map user content, network topology, and temporal dynamics across different topical domains.

Methodology: From Raw Data to Simulation-Ready Networks

The researchers developed an end-to-end pipeline to transform raw Reddit data into a structured input for simulations, moving through five key stages of refinement.

The end-to-end pipeline for transforming Reddit data into an empirically grounded dataset for agent-based modeling.
  1. Data Collection & Preprocessing: Posts and comments were gathered from Reddit across three distinct domains: Technology, Climate Change, and COVID-19. Rigorous filtering was applied to remove bots, spam, and deleted accounts.
  2. User Categorization (Aggregated Agents): Instead of modeling millions of individual users, the study used BERTopic and sentence-transformer clustering to group similar users into "aggregated agents." These agents are defined by their thematic focus, such as "ScienceTrust," "MaskUsage," or "GeneralChat."
  3. Inferred Social Ties: Since Reddit does not expose explicit "follow" relationships, the authors inferred directed social links based on temporal consistency. If User A replies to User B's posts across multiple time windows, the model records a persistent following relationship. This moves beyond simple interaction counts to capture sustained attention.
  4. Network Construction: A directed, weighted graph was constructed where nodes are agent categories and edges represent these inferred social connections, weighted by interaction frequency.
  5. Analysis and Benchmarking: The resulting networks were analyzed using standard metrics to provide a baseline for future simulations.

Aggregated Agents and Domain Diversity

The study introduces a new, publicly available dataset derived from Reddit, spanning approximately one million posts and comments. The dataset is organized into 33 technology-focused agents, 14 climate-focused agents, and 7 COVID-related agents.

Visualization of aggregated agents across the Technology, Climate, and COVID-19 domains, categorized by thematic focus.

This approach highlights how the dataset supports both traditional and LLM-driven "generative agents" by providing realistic conversational contexts. By grouping users into these thematic bubbles, researchers can simulate interactions between personas with established biases, interests, and interaction histories.

Key Insights: Topic-Dependent Social Structures

One of the most significant findings is that social structures are not "one size fits all." The topology of the interaction network changed drastically depending on the topic:

  • Climate Change: These discussions formed dense, highly clustered networks with high reciprocity (0.286) and a high clustering coefficient (0.765), suggesting sustained, multi-turn debates and a cohesive community.
  • COVID-19: Interactions were sparse and largely one-directional, reflecting an "episodic" or broadcast-style information sharing pattern with zero reciprocity at the inferred "following" level.
  • Technology: This domain exhibited the highest modularity (0.258) and the largest number of distinct communities (6), reflecting high topical specialization around specific technical niches.

Across all domains, a "GeneralChat" agent emerged as a central bridge. While specialized agents (like "QuantumComputing") stayed in their niches, these hub agents facilitated cross-topic flow, making them prime targets for intervention strategies in simulations.

Validation and Future Directions

A manual qualitative review of 105 interaction chains confirmed that the inferred agents followed realistic patterns of timing, sentiment progression (e.g., neutral to argumentative), and semantic coherence. This confirms that the "aggregated agent" approach captures genuine social tendencies, such as homophily—the tendency for users to interact with others who share similar emotional tones and writing styles.

While the model currently focuses on active replies and is rooted in Reddit's unique culture, future work aims to extend this methodology to other platforms and incorporate Dynamic Graph Neural Networks to track how these communities and "influencer" agents evolve in real-time.

Summary updated

Abstract

Agent-based modeling (ABM) provides a powerful framework for exploring how individual behaviors and interactions give rise to collective social dynamics. However, most ABMs rely on handcrafted or parameterized agent rules that are not empirically grounded, thereby limiting their realism and validation against observed data. To address this gap, we constructed a large-scale, empirically grounded dataset from Reddit to support the development and evaluation of agent-based social simulations. The dataset includes 33 technology-focused, 14 climate-focused, and 7 COVID-related aggregated agents, encompassing around one million posts and comments. Using publicly available posts and comments, we define agent categories based on content and interaction patterns, derive inter-agent relationships from temporal commenting behaviors, and build a directed, weighted network that reflects empirically observed user connections. The resulting dataset enables researchers to calibrate and benchmark agent behavior, network structure, and information diffusion processes against real social dynamics. Our quantitative analysis reveals clear topic-dependent differences in how users interact. Climate discussions show dense, highly connected networks with sustained engagement, COVID-related interactions are sparse and mostly one-directional, and technology discussions are organized around a small number of central hubs. Manual qualitative analysis further shows that agent interactions follow realistic patterns of timing, similarity between users, and sentiment change.
BibTeX
@misc{sittar2026constructing,
  author = {Abdul Sittar and Miha Češnovar and Alenka Guček and Jeff Z. Pan},
  title = {{Constructing a Dataset to Support Agent-Based Modeling of Online Interactions: Users, Topics, and Interaction Networks}},
  year = {2026},
  eprint = {2601.12628},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.12628},
}
Figure