Agent-based models capture how local decisions give rise to collective phenomena, but are rarely learnable from real data. This study introduces a differentiable framework that reconstructs individual behavioral rules while preserving interaction structure and stochasticity through the integration of graph neural networks and diffusion models. The approach bridges mechanistic modeling and machine learning, enabling empirical testing of theories about decentralized and emergent social and ecological systems.

Opinion dynamics models describe how social influence can lead groups toward consensus or conflict, but their empirical validation requires reliable parameter estimation. This work studies how to estimate the key parameters of a model of opinion formation. We prove that one parameter – the general "openness" – can be reconstructed given enough observations, while estimating how quickly opinions adjust is much more difficult, exhibiting persistent bias.

Estimating hidden opinions in agent-based models is challenging because individual beliefs are only partially observable. This paper compares two competing approaches for reconstructing those hidden states: Data Assimilation (DA) — widely used in applications like weather forecasting — and Likelihood-Based Inference (LBI), which directly exploits the model’s internal probabilistic structure but requires a manually designed likelihood function for each model. The study shows that LBI better recovers individual opinions, while DA remains effective for forecasting collective trends.

Understanding why people mobilize for collective action requires linking online behavior to social position, exposure, and ideology. Using longitudinal Reddit data and Bayesian causal modeling, this research disentangles how media attention, climate experiences, and peer dynamics jointly shape participation in climate activism. The results highlight how information diffusion and class-linked engagement transform awareness into sustained political mobilization.

Debates about echo chambers often assume that exposure across political lines fosters understanding, yet online interactions can reinforce ideological boundaries instead. Analyzing Reddit discussions on U.S. politics, this study examines how cross-party replies reshape engagement and community participation. The findings show that such encounters typically deepen within-group activity rather than bridging divides, revealing the fragile conditions under which opinion change and political integration may occur online.

Agent-based models of opinion formation often rely on repeated simulations to approximate observed collective outcomes, limiting both interpretability and efficiency. This work introduces a likelihood-based estimation approach that directly connects model parameters to data through probabilistic generative modeling, allowing opinions and interactions to be inferred from evidence rather than tuned by trial. By applying it to the bounded-confidence model, the study advances a data-driven understanding of how opinions evolve through social influence.

Agent-based models explain how micro-level rules produce macro-level patterns, yet linking them to data remains a challenge. This paper presents a protocol for inferring latent agent variables through probabilistic modeling and gradient-based expectation maximization, demonstrated on a housing market simulation. The method enhances predictive accuracy and interpretability, promoting a data-driven way to calibrate generative models of collective behavior.

Online debates are often portrayed as ideological echo chambers, yet the structure of conversation can mirror offline social boundaries instead. Analyzing millions of interactions in Reddit’s news discussions, this study shows that users connect more across ideological lines than across demographic divides such as age and income. The findings reveal how latent social stratification shapes digital communication, suggesting that online polarization often reproduces material divisions rather than belief systems.

Digital communities can replicate the social mechanisms of peer support traditionally found in offline recovery groups. This study analyzes interactions among Reddit users in opioid recovery, uncovering patterns of trust, reciprocity, and behavioral reinforcement that help sustain participation and progress. The results indicate that online spaces can act as functional surrogates for support networks, providing structure and belonging to individuals facing stigma and isolation.

Online discussions evolve through cascades of interactions that mirror collective opinion dynamics. This work introduces a probabilistic generative model that infers latent communities based on their degree of echo-chamber behavior and ideological polarity. By learning from both network structure and information propagation, the model identifies how polarized subgroups emerge and interact, providing a scalable statistical tool for understanding political and social fragmentation online.

Understanding ideology in social media requires models that learn from how information propagates, not just what it contains. This paper introduces a stochastic generative model that infers multidimensional ideological embeddings by fitting the flow of politically salient content across users. By capturing alignment through diffusion dynamics rather than labels, the model learns the structural complexity of political belief systems, offering a principled framework for representing ideology in networked communication.

Agent-based models explain how opinions evolve through interaction, yet they often remain detached from real data. This work develops a probabilistic inference framework that learns latent opinions and model parameters directly from social traces, retaining the interpretability of theoretical models while gaining empirical grounding. Applied to political conversations on Reddit, it provides a data-driven test of opinion dynamics theories, revealing limited evidence for the backfire effect in online debate.