Breaking the Limits: Synergizing LLM-Generated Feedback and Neighborhood Selection for Misinformation Detection

• GF-NSMD: Generated feedback and neighborhood selection to detect misinformation. • LLM generates diverse feedback to enrich data and improve misinformation modeling. • Neighborhood selection: Selects relevant neighbors, reducing noise. Social media has significantly increased the speed of information dissemination, but it has also expanded the reach of false content, posing challenges to public cognition and social governance. Existing graph-based detectors frequently rely on relatively dense connectivity and emphasize near-neighbor aggregation, which makes it difficult to recognize newly emerging misinformation in a timely and stable manner. To address these issues, we propose a novel model called Synergizing LLM-Generated Feedback and Neighborhood Selection for Misinformation Detection (GF-NSMD). GF-NSMD integrates diversified feedback generation with neighborhood selection to jointly enrich sparse social context and improve the utilization of informative long-range evidence in propagation graphs. The diversified feedback module uses a large language model to generate user comments and social-context explanations, enriching sparse contextual signals in imbalanced propagation structures. The neighborhood selection module then constructs k-hop propagation neighborhoods and adaptively highlights informative nodes, making the enriched long-range context more usable for classification. Experiments on six public datasets show that GF-NSMD achieves the most stable gains on fake news detection, improving macro f1-score by 1.6%-4.6% over state-of-the-art baselines. However, it does not consistently outperform competing methods on framing and propaganda detection tasks. In addition, GF-NSMD remains robust under limited-comment and noise-amplified settings, and ablation studies further validate the contributions of both the diversified feedback generation and neighborhood selection modules.