[Paper Review] Fake News Mitigation via Point Process Based Intervention

Mehrdad Farajtabar, Jiachen Yang+7 more|arXiv (Cornell University)|Mar 22, 2017

Misinformation and Its Impacts33 references99 citations

TL;DR

The paper proposes a multistage intervention framework that uses a multivariate Hawkes process to model fake news and mitigation events, integrating reinforcement learning to optimize interventions under budget constraints in real time. It demonstrates real-time Twitter experiments and synthetic-data superiority over baselines.

ABSTRACT

We propose the first multistage intervention framework that tackles fake news in social networks by combining reinforcement learning with a point process network activity model. The spread of fake news and mitigation events within the network is modeled by a multivariate Hawkes process with additional exogenous control terms. By choosing a feature representation of states, defining mitigation actions and constructing reward functions to measure the effectiveness of mitigation activities, we map the problem of fake news mitigation into the reinforcement learning framework. We develop a policy iteration method unique to the multivariate networked point process, with the goal of optimizing the actions for maximal total reward under budget constraints. Our method shows promising performance in real-time intervention experiments on a Twitter network to mitigate a surrogate fake news campaign, and outperforms alternatives on synthetic datasets.

Motivation & Objective

• Motivate the design of automated interventions to curb fake news in social networks without heavy manual oversight.
• Model fake news spread and mitigation as interacting point processes with exogenous control terms.
• Formulate fake news mitigation as a reinforcement learning problem with a networked MDP and budget constraints.
• Develop a policy iteration method tailored to multivariate networked point processes.
• Demonstrate real-time intervention capability and compare to baselines on synthetic data.

Proposed method

• Model fake news and mitigation as a multivariate Hawkes process with exogenous control terms.
• Define event exposure as BN(t) to quantify influence of fake and mitigation activities.
• Formulate the mitigation problem as a Markov decision process with state x and actions u, under budget constraints.
• Use a linear feature representation based on recent event counts to perform Least Squares Temporal Difference (LSTD) policy evaluation and policy improvement.
• Derive first and second order statistics for non-stationary Hawkes processes to enable reward evaluations.
• Implement a two-phase approach: offline value-function learning and online real-time intervention using a policy that selects exogenous intensities for mitigator nodes.

Experimental results

Research questions

• RQ1 How can fake news spread and mitigation activities be accurately modeled jointly in continuous time within a network?
• RQ2 Can we design an intervention policy that optimizes mitigation impact under budget constraints using point process dynamics?
• RQ3 What reward structures best capture mitigation effectiveness (e.g., correlation between mitigation and fake-news exposures or differential minimization of unmitigated fake events)?
• RQ4 Is a model-based, LSTD-based policy iteration approach effective for real-time interventions in Hawkes-process networks?
• RQ5 How do non-stationary, multivariate Hawkes statistics influence policy evaluation and improvement?

Key findings

• The framework enables real-time intervention experiments on a Twitter network and outperforms alternatives on synthetic datasets.
• Second-order statistics for non-stationary multivariate Hawkes processes are derived to support policy evaluation.
• A state representation combines past interval event counts to capture network dynamics for policy learning.
• LSTD-based policy evaluation with a model-based improvement step yields convergent policies under budget constraints.
• Reward function choices (correlation maximization and difference minimization) provide flexible objectives for mitigation effectiveness.