Multimodal learning with lifecycle awareness for early misinformation detection in social information systems

Abstract Misinformation often causes harm before rich propagation evidence becomes available, making early detection difficult for content-governance systems. This study proposes CIML-Net, a lifecycle-aware multimodal framework for early misinformation detection. The method divides an event into weak lifecycle stages using time, cumulative size, and growth-rate statistics, then encodes textual, visual, and event-structure signals with RoBERTa, CLIP, and a graph attention network. Stage-conditioned affine fusion and cross-stage contrastive learning learn representations that remain stable as an event evolves. Experiments use event-disjoint evaluation on Weibo21 and PHEME, with CHEF serving as text-only transfer validation, and compare CIML-Net with propagation-only, multimodal, and graph-contrastive baselines as well as three recent journal methods published in 2025–2026. CIML-Net improves early-window detection over all of them: on Weibo21 it achieves 80.1% Early-F1@6h and 91.0% full-trajectory F1, exceeding the strongest recent baseline by 2.4 points in the 6 h window, and mechanism-matched ablations link the early-window gain to the cross-stage positive definition under fixed encoder capacity. Event-cluster bootstrap intervals exclude zero for the principal early-window deltas, and a nine-fold leave-one-macro-event-out evaluation on PHEME preserves the method's aggregate advantage under unseen macro events. Results indicate that lifecycle-conditioned multimodal learning improves early-window detection under the evaluated benchmarks and a bounded single-node serving setting.