Introduction
Understanding the underlying motivations of actors in social networks is crucial for combating misinformation. While game-theoretic models simulate interactions between malicious sources and defenders, they often assume that cost functions—the strategic goals of players—are known. In reality, these motivations are latent. Inverse Reinforcement Learning (IRL) offers a path to infer these functions, but traditional IRL models struggle with the complexities of human behavior, specifically memory horizons and cognitive biases.
Problem Formulation
Standard IRL typically relies on Markov Decision Processes (MDPs), assuming the next state depends only on the current one. However, human information consumption is influenced by a "memory horizon." Furthermore, humans exhibit cognitive asymmetry: confirmation bias leads us to favor information aligned with our existing beliefs, while novelty bias draws us toward outlying content. Finally, social environments are dynamic; cost functions shift in response to real-world events like elections or pandemics.
Methodology: The $M^3IRL$ Framework
To address these challenges, we introduce $M^3IRL$ (Memorized Model and Maximum-entropy based Inverse Reinforcement Learning). The framework consists of two primary components: a social information-diffusion model and a model-based IRL algorithm.
The diffusion model characterizes an individual's opinion at time $k+1$ as a function of their subconscious bias and a memory of surrounding opinions over a horizon $\tau$. Influence weights are mathematically tuned to capture asymmetric biases. The $M^3IRL$ algorithm then applies the principle of Maximum Entropy to handle behavioral uncertainty. Unlike standard IRL, it operates in a non-MDP framework, requires only a single trajectory of data, and manages bounded opinion variables within the $[-1, 1]$ range.
Cognitive Behavioral Asymmetry
A core innovation of this work is the formalization of asymmetric influence patterns. The "distance" between two opinions is perceived differently based on the observer's current stance.
As shown in Figure 6, opinions closer to a user's current belief exert a disproportionately larger influence (confirmation bias), while information sources may strategically use "outlying" content to leverage novelty bias and grab attention. By incorporating decaying memory influence based on the ACT-R cognitive architecture, the model reflects how older information loses impact over time.
Experimental Results
The model was validated using a Twitter COVID-19 vaccine dataset from 2021. Tweets were processed via a BERT model to map sentiments onto a scale from -1 (opposing) to +1 (supporting). In a network of 9 users, $M^3IRL$ successfully identified the opposing goals of two information sources: one steering public opinion against vaccines and another supporting them. Notably, the algorithm quantified influence, finding that the pro-vaccine source had a significantly higher influence weight ($\alpha = 0.7960$) than the anti-vaccine source ($\alpha = 0.2040$). Analysis across different time windows confirmed that the model could capture evolving strategies in response to external events.
Conclusion
$M^3IRL$ provides a robust framework for uncovering hidden strategic goals in social networks by accounting for human memory and cognitive biases. While current validation is limited to small-scale networks, future work will focus on computational optimizations for large-scale application and the integration of automated detection to distinguish between human and bot-like behaviors.
Summary updated