When a new medical claim spreads online, such as a miracle cure or a doctored lab result, there are almost no labeled examples of it to learn from. Detectors trained on older rumors tend to memorize the stories they saw rather than anything general about deception. This proposal asks whether the shape and timing of a spreading conversation can fill that gap. A shape signal would be who replies to whom, how deep the reply tree grows, and when it suddenly surges. Unlike the text, this signal should carry over from past rumors to new ones, because deceptive and debunked claims leave recognizable geometric traces regardless of topic.
Concretely, we pretrain an encoder on unlabeled conversation cascades by teaching it to predict how each cascade grows over time: which existing post the next reply attaches to, and how long until it arrives. We then attach a small classifier trained on a few dozen labeled health cascades and fuse it with the claim's text and any attached clinical image. A learned gate trusts content less when the claim is new. The central bet is that temporal growth structure is the most label-efficient and topic-robust signal available. We will test that bet against a strong simple competitor: hand-counted tree statistics fed to logistic regression.
Core Claim
Under evaluation splits that keep events apart and move forward in time, with 16 to 128 labeled health-claim cascades, an encoder pretrained to predict cascade growth will beat three alternatives in both accuracy and calibration: supervised graph networks, language models pretrained on static reply structure, and hand-crafted structural features. The advantage should be largest in the first hours of spread, when content cues are thin and the growth dynamics are the main evidence.
Several findings motivate this claim. Rumor models trained on static datasets overfit to surface wording and effectively memorize stories, as shown by Examining the Limitations of Computational Rumor Detection Models Trained on Static Datasets. That paper is also why our evaluation must be event-disjoint and time-forward.
The diffusion literature documents topic-independent shape signatures:
- Deep branching versus shallow broadcast. Misinformation tends to form deep, multi-generation trees, while authentic news forms shallow star broadcasts (Zhao et al.; Liu).
- Debunking stars. Debunked claims attract dense, shallow stars of independent refutations (Shang et al.).
- Two-stage surges. Rumors often simmer among low-influence accounts until a central account amplifies them (Ma et al.).
- Narrow chains in emergencies. During crises, spread degenerates into narrow chains (Study on the rumor detection of social media in disaster based on multi-feature fusion method).
The last two signatures are temporal: they concern the order and speed at which the graph appears, not just its final shape. That is the spatio-temporal element this proposal treats as the core mechanism.
Relation to Prior Work
The method builds on and differs from four lines of work.
Bidirectional graph convolution. Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks encodes the final reply tree in both directions, feeds the root claim into every layer, and regularizes by randomly dropping edges. It is fully supervised and sees a static snapshot. We reuse its bidirectional encoder as our backbone but replace supervised training with growth pretraining and feed it time-stamped edges.
Static structural pretraining. Enhancing Rumor Detection Methods with Propagation Structure Infused Language Model continues pretraining a language model on unlabeled cascades with three static tasks: is this post a root reply, are two posts in the same branch, is one the parent of another. Our pretraining is the closest analogue and the one we must beat directly. We add the order and timing of attachment, which their tasks discard.
Zero-shot prompting. Zero-Shot Rumor Detection with Propagation Structure via Prompt Learning targets the no-label regime by encoding tree depth and parent offsets into prompts. It supplies both a low-label baseline and evidence that structure transfers across domains. It uses breadth-first ordering rather than real arrival times.
Multimodal hierarchy. Exploring Hyperbolic Hierarchical Structure for Multimodal Rumor Detection combines images with tree hierarchy in hyperbolic space, which suits tree-shaped data. It is the nearest multimodal competitor. Its hierarchy is static, and it is not designed for the few-label clinical setting.
The clinical gap. The research file contains no paper specifically on clinical-image or medical-claim rumor detection. The "clinical" modality here is an application regime we define and annotate ourselves, not an established benchmark.
Research Program
Pretraining objective. Treat each cascade as a sequence of arrivals $(p_i, t_i)$, where $p_i$ is the existing post that reply $i$ attaches to and $t_i$ is its timestamp. Let $G_{<i}$ be the time-stamped tree before arrival $i$. The encoder is pretrained on unlabeled cascades to minimize
$$ \mathcal{L} = -\sum_i \log P_\theta(p_i \mid G_{<i}) \;-\; \lambda \sum_i \log f_\theta(t_i - t_{i-1} \mid G_{<i}), $$
where:
- $P_\theta$ is a softmax over current nodes, scored by the bidirectional graph encoder.
- $f_\theta$ is a log-normal density over waiting times.
- $\lambda$ balances the two terms.
The attachment term forces the encoder to represent where the tree is branching, which captures depth and star shape. The timing term forces it to represent surges and stalls.
Fine-tuning with few labels. A linear classifier on the pooled representation is trained with 16, 32, 64, or 128 labeled cascades. Claim text and any attached image, such as a lab report, scan, or chart, are encoded by frozen pretrained encoders. A scalar gate conditioned on the claim's embedding-space distance to training claims decides how much to weight content versus structure. The intent is that novel claims lean on structure, since content is exactly what the static-dataset critique says fails to transfer.
Datasets. The Weibo benchmark used by Bi-GCN is the primary structural dataset. From it we construct two things:
- A health-topic subset.
- Event-disjoint, time-forward splits, in which all training cascades predate all test cascades and no event spans both.
We will also annotate a small held-out set of medical claims with attached clinical imagery. We have no reported results for that set, and the research file reports no numbers on any health-specific split.
Baselines
Do-nothing and content-only references. The naive floor is the majority class. The content-only reference is a text classifier (plus image encoder where images exist) fine-tuned on the same labeled cascades. Given the static-dataset findings, we expect it to degrade sharply under time-forward splits. The gap between random and time-forward splits for this model measures how much of any reported number is story memorization.
Bi-GCN. The strongest supervised structural method is Bi-GCN (Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks), which reports 96.1% accuracy on Weibo under its original split. We will rerun it on our splits with the same labeled budget, the same text features, and a matched hyperparameter search of equal trial count. We will report both its original-split and time-forward numbers. Beating it means higher macro-F1 on time-forward, event-disjoint splits at 64 labels or fewer, averaged over at least five seeds with non-overlapping confidence intervals. We do not expect to beat 96.1% on the original split, and that is not the target.
Static structural pretraining. The PEP-pretrained language model (Enhancing Rumor Detection Methods with Propagation Structure Infused Language Model) is the most direct competitor on the pretraining question. We will pretrain it on the identical unlabeled cascade pool with a matched compute budget.
Zero-shot and multimodal methods. Response-aware prompt learning (Zero-Shot Rumor Detection with Propagation Structure via Prompt Learning) serves as the zero-label reference. RumorCone (Exploring Hyperbolic Hierarchical Structure for Multimodal Rumor Detection) serves as the multimodal reference on the image subset. We found no reported result on Lacuna for either method on health-specific or time-forward splits.
Hand-crafted structural features. This is logistic regression on roughly twenty counted features: maximum depth, breadth, star ratio, fragmentation, time to first surge, and the fraction of replies arriving in the first hour.
Ablations. Three ablations isolate the key ingredient:
- The same encoder pretrained with only the attachment term ($\lambda=0$).
- The same encoder pretrained with PEP's static tasks.
- The full model with the content gate fixed at equal weights.
Why not just…?
Why not just use hand-counted tree statistics with logistic regression? This is the real threat. The signatures in the literature, such as depth, star shape, and surge timing, are exactly what such features measure, and with 32 labels a twenty-feature linear model is hard to beat. We will run it at every label budget. If it matches the pretrained encoder within confidence intervals, the proposal's claim fails, and we will report that the transferable signal is fully captured by known descriptors. To give the learned model a fair chance at adding value, we also test a hybrid that feeds these features alongside the learned representation.
Why not just use the static-structure pretraining that already exists? PEP already shows that unlabeled cascades help. Its three tasks, however, see only the final tree, so they cannot represent the two-stage surge that Ma et al. identifies. The matched-compute comparison and the PEP-task ablation on our own encoder settle whether time adds anything beyond static structure.
Why not just train Bi-GCN with more regularization? Bi-GCN's edge-dropping already targets overfitting. Its 96.1% was obtained with full supervision on a split where related stories can appear in both train and test. The learning curves and time-forward splits show directly whether its accuracy survives when labels are scarce and stories are new.
Why not just classify the claim text with a large pretrained model? The static-dataset study shows that text-trained detectors memorize stories. Novel medical claims are exactly the case where memorized stories do not apply. The content-only baseline under time-forward splits quantifies this, and the learned gate lets us test whether content helps once structure is present.
Proposed Experiments and Decisive Checks
We will measure four things:
- Learning curves: macro-F1 and area under the ROC curve at 16, 32, 64, and 128 labels.
- Calibration: expected calibration error.
- Early detection: performance when only the first 1, 3, 6, and 12 hours of each cascade are visible.
- Clinical transfer: all of the above on the held-out clinical-image set.
The decisive check has three conditions:
- The growth-pretrained encoder beats both the hand-crafted-feature model and the PEP-pretrained model at 32 and 64 labels on time-forward splits.
- Removing the timing term erases most of that margin.
- The margin is largest in the earliest time windows.
If the first condition fails against the hand-crafted features, we conclude that the transferable structure is already known and small enough to count. That finding would still be useful for clinical deployment, where a transparent twenty-feature model is easier to audit than a pretrained graph encoder.
References
- Examining the Limitations of Computational Rumor Detection Models Trained on Static Datasets
- Fake news propagates differently from real news even at early stages of spreading
- The Dynamics of Misinformation Spread on Social Media Networks
- FauxWard: a graph neural network approach to fauxtography detection using social media comments
- Detect Rumors in Microblog Posts Using Propagation Structure via Kernel Learning
- Study on the rumor detection of social media in disaster based on multi-feature fusion method
- Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks
- Enhancing Rumor Detection Methods with Propagation Structure Infused Language Model
- Zero-Shot Rumor Detection with Propagation Structure via Prompt Learning
- Exploring Hyperbolic Hierarchical Structure for Multimodal Rumor Detection