An interpretable multi-scale temporal-structural feature framework for early rumor detection on social media

It is important to identify which propagation characteristics help distinguish rumors from verified or non-rumor information when developing transparent detection systems. We introduce MSTRD, an explainable multi-scale temporal-structural feature framework for rumor detection on social media. MSTRD represents each propagation tree using a 78-dimensional feature vector organized into structural topology (35 features), temporal dynamics (28 features), and cascade shape features (15 features). Classification is performed using a soft-voting ensemble of HistGradientBoosting, ExtraTrees, and a multi-layer perceptron. Experiments on three standard benchmarks yield four principal findings: (i) MSTRD achieves 93.5 ± 0.4% accuracy on Weibo binary classification, a value that falls within the range reported in the literature for end-to-end graph models on this benchmark (PAGCN 93.9% [1]; RvNN 90.8% [2]); because those models learn their own node representations while MSTRD applies fixed hand-designed descriptors, these published figures are cited as reference points rather than as a controlled head-to-head comparison, and the result is offered only as evidence that engineered propagation features retain measurable discriminatory value for binary rumor detection; (ii) temporal features are the most informative feature group, with a 2.4%-point accuracy drop when removed; (iii) false rumors are consistently easier to identify than true rumors, with precision of 0.901 on Twitter15 and 0.926 on Twitter16; and (iv) the framework achieves 91.4% early detection accuracy on Weibo within 1 h of the initial post. Feature-importance analysis indicates that mean tree depth (28.9%) and Wiener index approximation (26.9%) jointly account for 55.8% of measured importance, providing hypothesis-generating design guidance for future depth-aware GNN architectures, subject to direct validation.