Advancing rumor detection in social media: A quality-stratified systematic review from CNNs to LLM-GNN hybrids

Rumor detection on social media is imperative for countering misinformation amid the post-pandemic proliferation of multimodal content. However, prior surveys (2021–2023) inadequately address large language model integration, methodological inconsistencies inflating performance claims, and propagation-aware hybrids’ deployment challenges, underscoring the need for an updated, evidence-stratified synthesis incorporating 2024–2025 advancements. This systematic review evaluates 77 deep learning-based studies from 2021 to 2025 via a novel S1–S8 quality rubric emphasizing reproducibility and generalizability, categorizing features into text-based (lexical/syntactic/semantic embeddings), visual (tampering/multimodal fusion), user-centric (profiles/behaviors), and propagation-based (graphs/temporals) domains, while synthesizing deep learning techniques such as convolutional neural networks for local patterns, recurrent neural networks for sequences, graph neural networks for relations, and large language models for semantics. Meta-analysis, stratified by post- versus event-level units and full- versus early-context modes, demonstrates that hybrid graph neural network-large language model architectures surpass standalone models with 5%–15% F1-score gains on benchmarks like PHEME, Twitter15/16, and Weibo (see Section 3.6 and Supp. C Table 2 for detailed aggregation) highlighting trends in zero-shot adaptation, joint stance-emotion learning, and explainable mechanisms. Enduring challenges include bot distortions, early-detection fragility, adversarial vulnerabilities, cross-lingual shifts, data imbalances, ethical biases, and multimodal scalability, amenable to mitigations via temporal graph neural networks and contrastive alignment. Future priorities encompass standardized multilingual datasets, human-in-the-loop interpretability, real-time federated learning, and ethical auditing, furnishing researchers with weighted insights for resilient, ethical rumor detection in heterogeneous digital ecosystems.