Causal Influences over Social Learning Networks

Problem Statement
Identifying influential agents within social networks is a critical challenge with applications ranging from detecting propaganda to optimizing targeted advertising. However, traditional influence estimation methods often rely on network topology metrics (e.g., eigenvector centrality) or descriptive importance factors, which fail to distinguish between correlation and causation. In social networks, information spreads and mixes over time, creating ripple effects and confounding factors (such as homophily or shared external sources) that obscure true causal relationships. Furthermore, existing causal approaches often rely on randomized experiments, which are frequently infeasible in real-world social settings, or utilize predictive tools like Granger causality that are not robust against latent confounders. This paper addresses the need for a rigorous framework to quantify causal influences between agents in dynamic social learning networks using observational data.

Methodology
The authors approach the problem through the lens of Structural Causal Models (SCMs), defining influence as the change in an agent's behavior resulting from an intervention. The study utilizes two established mathematical models for social learning over graphs:

1. Non-Bayesian Social Learning (NBSL): A framework where agents update beliefs via a local Bayesian step followed by a geometric averaging of neighbors' beliefs.
2. Adaptive Social Learning (ASL): An extension that introduces temporal weighting parameters to prioritize recent observations, allowing for adaptation in non-stationary environments.

The core methodology involves analyzing the "atomic and persistent intervention" on a specific agent $m$, where $m$'s belief is fixed to a constant value ($do(\mu_m)$). The authors derive closed-form expressions for the resulting asymptotic log-belief ratios of all other agents in the network. These expressions depend on the network topology (represented by the combination matrix $A$) and the agents' individual informativeness (quantified by Kullback-Leibler divergences).

To handle real-world scenarios where model parameters are unknown, the paper proposes a Graph Causality Learning (GCL) algorithm. This method estimates the necessary parameters (combination weights and agent informativeness) directly from raw observational data, specifically sequences of shared intermediate beliefs (actions) and the known graph topology (adjacency matrix).

Key Contributions

• Causal Impact Criterion: The paper proposes a novel criterion to quantify the causal impact of one agent on another for every pair $(m, k)$ in a dynamic network. This is derived for both NBSL and ASL frameworks, providing closed-form expressions that reveal how instantaneous effects diffuse over time and space.
• CausalRank Algorithm: To rank agents by their overall influence, the authors introduce CausalRank. Unlike simple averaging or standard centrality measures, this algorithm constructs a bipartite causal influence matrix and computes its principal eigenvector. This approach weights the influence of an agent based on the influence of the agents they affect, effectively identifying agents who are influential on other influential agents.
• Graph Causality Learning (GCL): An algorithm is introduced to estimate causal influences from observational data without requiring ground truth on model parameters. The paper provides a theoretical bound showing that the estimation error decreases as $O(1/\sqrt{M})$ with the number of samples $M$.
• Analysis of Special Cases: The study explicitly analyzes how causal effects depend on network parameters in specific topologies, such as fully-connected (federated) architectures and unidirectional rings, illustrating the roles of network centrality, agent informativeness, and forgetting factors.

Results

• Synthetic Data: Simulations on an 11-agent network demonstrate that the causal influence matrix differs significantly from the raw combination matrix. For instance, agents with low informativeness are shown to be more susceptible to influence, and indirect influence paths can be stronger than direct ones due to information mixing.
• Robustness: The CausalRank algorithm is shown to be robust against "follower-inflation" attacks (where bot accounts artificially inflate a user's follower count). While traditional metrics like follower count and average influence ranking (AIR) are sensitive to such manipulation, CausalRank maintains a stable ranking because it accounts for the actual causal impact rather than just the number of connections.
• Real-World Application: The GCL algorithm was applied to a sub-network of 20 Twitter users discussing cryptocurrency. Using natural language processing to extract sentiment from tweets as belief inputs, the study identified Elon Musk (User 1) as the most influential agent. Notably, the causal ranking differed from eigenvector centrality; for example, User 2 had high centrality but low causal impact, likely due to low informativeness (e.g., retweeting without adding new information).
• Correlation vs. Causation: Experiments confirmed that while increasing the correlation of external observations changes the correlation of agents' beliefs, the estimated causal effects remain constant, demonstrating the method's ability to filter out non-causal confounding factors.

Significance
The paper claims to be the first study to examine the diffusion of influence over space and time in social learning networks from a causal inference perspective. By treating influence as a causal quantity derived from structural causal models, the work provides a method to discard spurious associations that plague correlation-based approaches. The proposed framework is significant for both modeling opinion formation in human social networks and designing distributed decision-making systems for multi-agent systems (e.g., sensors, robots). The ability to learn causal structures from observational data makes the approach particularly valuable for real-world social media analysis where controlled experiments are impractical. The authors modestly note that while their results are robust, the lack of ground truth in real-world social networks remains a challenge, and they leave the exploration of Large Language Model (LLM) based controlled experiments for future work.