New Bandit Algorithm Balances Peer Influence Estimation and Rewards

In networked environments, understanding who influences whom is critical for viral marketing, information diffusion, and recommendation systems. Traditional methods often rely on static observational data that can only capture correlations, not causal influence. Online learning approaches can estimate influence probabilities through interventions, but they struggle with a fundamental dilemma: focusing on minimizing regret (i.e., picking the best recommendation) biases exploration toward high-influence regions, while focusing on reducing estimation error requires costly uniform exploration.

To solve this, the authors model peer influence as a contextual linear bandit problem where each recommendation is a context-dependent decision. They formally characterize the achievable rate pairs between regret and mean squared estimation error, proving that no algorithm can simultaneously achieve optimal rates for both. They then propose an uncertainty-guided exploration algorithm that interpolates between the two extremes via a tunable parameter. Experiments on semi-synthetic networks (e.g., Facebook, Wikipedia) demonstrate that the method outperforms both static baselines and standard contextual bandits, enabling practitioners to tailor the trade-off to their specific needs—whether maximizing campaign ROI or precisely mapping influence paths.