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

A new study reports that a transparent, feature-based framework can detect rumors on social media with accuracy close to that of far more complex graph neural networks on a binary benchmark, while revealing exactly which propagation characteristics drive the decision. The framework, called MSTRD, represents each rumor's spread as a propagation tree and converts it into a 78-dimensional vector of hand-designed descriptors, then classifies it with a soft-voting ensemble of three conventional machine-learning models. On the Chinese platform Weibo, the system reached 93.5 ± 0.4% accuracy under 5-fold stratified cross-validation, a figure the author carefully positions within the range reported for end-to-end graph models on the same benchmark, including 93.9% for PAGCN and 90.8% for the earlier RvNN.

The study, published in Discover Artificial Intelligence, was motivated by an interpretability gap. Graph neural networks such as BiGCN, GCAN, EBGCN, GACL and PAGCN learn their own node representations directly from propagation graphs and achieve strong benchmark results, but they offer limited direct evidence about which structural and temporal properties of rumor spreading are actually diagnostic. Without that evidence, it is difficult both to explain model behaviour and to identify where architectural improvements should be made.

MSTRD was designed not as another end-to-end architecture but as a complementary diagnostic tool that systematically extracts multi-scale temporal and structural features and measures which feature families carry the clearest discriminative signal. The 78 features are organized into three groups that can be ablated independently. The structural group, with 35 dimensions, captures tree size, leaf counts and ratios, maximum, average, median and standard-deviation depths, degree distribution statistics, branching factors, the fraction of nodes at each of eight tree levels, a Wiener index approximation of mean path length, subtree size statistics and degree entropy.

The temporal group, with 28 dimensions, encodes cascade duration, inter-event interval statistics and percentiles, four-stage propagation speeds based on quartiles, a burstiness coefficient capturing irregularity in posting patterns, time-to-fraction milestones at 10% through 90% of final cascade size, a Gini coefficient of activity concentration and first-response time. A cascade-shape group of 15 features adds width-to-depth ratios, shape asymmetry, a structural virality index following Goel and colleagues, star-likeness and depth-balance ratios and three cross-domain interaction terms. Classification is performed by three complementary base classifiers whose predicted probabilities are averaged.

HistGradientBoosting learns threshold-based boundaries and provides impurity-based feature rankings; ExtraTrees, with 150 randomized trees, aggregates robustly and captures multi-feature interactions while resisting overfitting; and a two-layer multi-layer perceptron models non-linear interactions unavailable to tree methods. All features are standardized before classification, and the ensemble produces a two-class output on Weibo and a four-class output on the Twitter benchmarks, distinguishing non-rumor, true rumor, false rumor and unverified categories. Evaluation used three standard benchmarks: 4,664 binary-labeled Weibo events, 1,490 four-class Twitter15 source tweets and 818 four-class Twitter16 source tweets, all under 5-fold stratified cross-validation.

The headline Weibo result comes with an important caveat that the author states explicitly. Because PAGCN learns node representations end-to-end while MSTRD applies fixed hand-designed descriptors, and because the two figures come from different implementations, splits and training protocols, the numerical proximity is reported as evidence that engineered propagation descriptors capture a comparable amount of binary-discriminative signal on this benchmark, not as a matched comparison or a claim of parity. On the harder four-class Twitter tasks, the framework reached 72.6% accuracy on Twitter15 and 71.6% on Twitter16, more than 15 percentage points below several recent graph-based models.

The author concludes that the representation is useful for diagnostic analysis and binary detection but is not sufficient as a standalone method for fine-grained rumor subtype prediction. Per-class analysis revealed a consistent asymmetry in detection difficulty. False rumors were the easiest category to identify, with precision of 0.901 on Twitter15 and 0.926 on Twitter16 and F1-scores of 0.878 and 0.891 respectively, while true rumors were consistently the hardest, with F1-scores of 0.592 and 0.607.

The author links this pattern to distinct propagation morphologies: false rumors tend to form deep, narrow trees with a mean depth of 5.8 and high burstiness of 2.1, whereas non-rumors spread in shallow, wide trees with a mean depth of 2.4 and gradual temporal spread. This echoes earlier findings that false rumors spread faster and more broadly than the truth, and it implies that systems aimed specifically at flagging false rumors can be expected to achieve substantially higher reliability. Early detection results address a key deployment requirement: identifying rumors before they reach a wide audience.

The model was retrained independently for each observation window from 1 hour to 24 hours after the initial post. On Weibo, accuracy reached 91.4% within 1 hour, only 2.1 percentage points below the 24-hour value of 93.5%. Few responses arrived between 1 and 2 hours and most cascades were structurally unchanged during that interval, which explains the identical accuracies at those two deadlines.

The steepest accuracy improvements occurred in the 2-to-4-hour window. On Twitter16, performance peaked at 75.0 ± 2.8% at 8 hours and declined to 71.3 ± 3.1% at 12 hours; because the dispersion intervals overlap, the author interprets this non-monotonic dip cautiously, noting that the 3.7-point difference corresponds to roughly six events per fold in the smallest benchmark and could reflect ordinary resampling variation rather than genuine degradation. Ablation experiments on Weibo quantified each feature family's contribution.

Removing the temporal features caused the largest drop, 2.4 percentage points, confirming that multi-scale temporal dynamics are the most informative single group for binary detection. Structural features contributed a more modest 0.4 points when removed. Removing the cascade-shape group actually raised Weibo accuracy by 0.6 points, indicating redundancy there, yet the same removal lowered four-class Twitter accuracy by 0.8 points, showing that redundancy is a property of a feature group relative to a task rather than an intrinsic property of the features.

A refined variant, MSTRD-R, retained nine non-redundant cascade-shape features and achieved 94.0% on Weibo, though the author notes this is indistinguishable from fold-level dispersion and that the small Twitter gains should be read as simplification evidence rather than meaningful improvement. Feature-importance analysis from the HistGradientBoosting pathway identified mean tree depth as the dominant signal at 28.9% of measured importance, followed by the Wiener index approximation at 26.9%, so that these two descriptors jointly account for 55.8%. Degree-distribution features and a high inter-event percentile followed.

Instance-level SHAP case studies illustrated how individual decisions arise: a correctly classified false rumor showed strong positive contributions from mean depth and burstiness, a non-rumor showed negative depth contributions and positive width contributions, and a misclassified true rumor revealed competing structural and temporal signals. Extracted decision rules, such as mean depth above 4.2 combined with a burstiness coefficient above 1.8 implying false rumor with 0.91 precision, achieved high precision but low recall and limited coverage, so the author recommends them as audit rules rather than a deployable detector. The study also surfaced a cautionary result about cascade size.

On a held-out Weibo fold, very large cascades of more than 50 nodes were 94.7% rumors, and within that bin the model reached 99.5% accuracy but a macro-F1 of only 49.9%, meaning it had effectively learned to emit the majority label while retaining almost no ability to recognize the rare non-rumor. The author emphasizes that size-stratified sampling, which raised macro-F1 to 78.3%, is a diagnostic instrument rather than a deployable remedy, since the cascade-size distribution in deployment cannot be controlled; model-side mitigation such as cost-sensitive training or size-conditional thresholds is left to future work. The paper closes with hypothesis-generating design suggestions for future graph architectures, including depth-aware aggregation and explicit burstiness encoding, all of which the author stresses require direct validation, and lists limitations ranging from the absence of textual content to the use of offline benchmarks rather than real-time streaming deployment.

Subject of Research: Interpretable multi-scale temporal and structural feature analysis for early rumor detection on social media propagation trees

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

Article References: Dai, N. (2026). An interpretable multi-scale temporal-structural feature framework for early rumor detection on social media. Discover Artificial Intelligence, 6(1), Article 974. https://doi.org/10.1007/s44163-026-02080-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02080-8

Keywords: rumor detection, social media, propagation trees, interpretability, graph neural networks, feature engineering, early detection, Weibo, Twitter15, Twitter16, machine learning, misinformation