Rumor Adaptive Graph Contrastive Learning

• The paper proposes centrality-guided adaptive augmentation to emphasize informative deep substructures in rumor propagation trees.
• RAGCL improves classification accuracy by tailoring node and edge perturbations based on the statistical observation of wide, shallow reply patterns.
• It combines supervised and contrastive objectives within a graph convolutional network to robustly address standard and low-resource rumor detection.

Searching arXiv for the specified papers to ground the article and confirm metadata. Rumor Adaptive Graph Contrastive Learning (RAGCL) denotes a graph-based contrastive learning approach for rumor detection on social media in which rumor propagation structures are encoded as graphs and augmented to improve graph-level classification. In the formulation introduced by "Propagation Tree Is Not Deep: Adaptive Graph Contrastive Learning Approach for Rumor Detection" (Cui et al., 10 Aug 2025), RAGCL is defined around rumor propagation trees (RPTs), adaptive view augmentation guided by node centralities, and a joint supervised-plus-contrastive objective tailored to the empirical observation that real-world RPTs are typically wide rather than deep. The term also appears in the provided technical description associated with "A Unified Contrastive Transfer Framework with Propagation Structure for Boosting Low-Resource Rumor Detection" (Lin et al., 2023), where it denotes a contrastive transfer framework for low-resource and cross-lingual rumor detection. Taken together, the literature in the provided record presents RAGCL as a family of contrastive graph-learning formulations for rumor detection, with the 2025 work (Cui et al., 10 Aug 2025) giving the specific expansion "Rumor Adaptive Graph Contrastive Learning."

1. Conceptual scope and problem setting

In (Cui et al., 10 Aug 2025), the basic learning problem is graph-level rumor classification over a labeled dataset $\mathbb{C}=\{c_i\}_{i=1}^m$, where each claim $c=(y,G)$ consists of a ground-truth label and a propagation tree $G=(V,E)$. The label space is either binary, $y\in\{N,R\}$, or fine-grained, $y\in\{N,F,T,U\}$, and the graph contains a source post $v_0$ together with comments linked by directed “reply-to” edges. The model is trained as a graph-level classifier

$$f: \mathbb{G}=\{G_i\}_{i=1}^m \longrightarrow \{y_i\}_{i=1}^m.$$

A central premise of (Cui et al., 10 Aug 2025) is that most existing graph-based models presume rumor propagation trees have deep structures and learn sequential stance features along branches, whereas statistical analysis on real-world datasets indicates that RPTs exhibit wide structures, with most nodes being shallow 1-level replies. On that basis, RAGCL is designed to focus learning on intensive substructures through adaptive augmentation rather than treating all parts of the tree uniformly. This suggests a shift from branch-sequence modeling toward importance-weighted structural perturbation.

The formulation associated with (Lin et al., 2023) addresses a related but distinct setting: low-resource rumor detection under domain and language shift. There, a rumor “event” is represented as an undirected propagation graph $G=(V,E)$ whose nodes are the root claim and all replies, with post-level features obtained from XLM-RoBERTa and propagated through a Multi-scale Graph Convolutional Network. In that setting, the contrastive objective is explicitly transfer-oriented, aligning source and target domains while applying target-wise contrastive regularization to few-shot target events (Lin et al., 2023).

2. Structural assumptions about rumor propagation

The defining empirical claim of (Cui et al., 10 Aug 2025) is that rumor propagation trees are very “wide” and that only a small fraction form deep subtrees. This observation motivates three augmentation principles intended to preserve informative parts of the graph while allowing stochastic perturbation elsewhere.

The first principle is Exempt Root: the root node $v_0$ is never dropped or masked. The second is Retain Deep-Reply Nodes: nodes and edges that anchor deeper subtrees should have lower drop probabilities. The third is Preserve Lower-Level Nodes in Deep Sections: within a deep chain, nodes closer to the root are more important than their deep successors (Cui et al., 10 Aug 2025).

These principles are significant because they explicitly encode a structural bias derived from dataset statistics rather than relying on generic graph augmentation heuristics. In the terminology of (Cui et al., 10 Aug 2025), adaptive view augmentation is guided by node centralities so that representation learning concentrates on rare but informative deep substructures while still emphasizing the root. A plausible implication is that the method is not merely contrastive learning on trees, but contrastive learning constrained by an empirically grounded model of how rumor conversations are distributed across reply depths.

By contrast, the propagation structure in (Lin et al., 2023) is modeled as an undirected topology “for enhancing the interaction of user opinions,” and the augmentations are event-level rather than node-centrality-driven. The juxtaposition of (Cui et al., 10 Aug 2025) and (Lin et al., 2023) therefore exposes two different structural assumptions: top-down, depth-aware tree adaptation in the former, and undirected cross-event transfer learning in the latter.

3. Centrality-guided adaptive augmentation

In (Cui et al., 10 Aug 2025), RAGCL uses standard centrality measures $\varphi_c:V\to\mathbb{R}^+$ to quantify node importance. The three measures specified are degree centrality,

$c=(y,G)$0

betweenness centrality,

$c=(y,G)$1

and PageRank centrality,

$c=(y,G)$2

with PageRank computed on the bottom-up graph. Node and edge importance scores are then defined as

$c=(y,G)$3

RAGCL converts these importance scores into drop or mask probabilities by normalizing in log-space. For nodes,

$c=(y,G)$4

and, given a global node-drop hyperparameter $c=(y,G)$5, each non-root node is dropped with probability

$c=(y,G)$6

while the root receives $c=(y,G)$7. Attribute masking uses the same probability $c=(y,G)$8, setting $c=(y,G)$9 when masking occurs. For edges,

$G=(V,E)$0

and, with edge-drop hyperparameter $G=(V,E)$1,

$G=(V,E)$2

The paper specifies node dropping, attribute masking, and edge dropping as the augmentation operators. It also notes that the edge $G=(V,E)$3 from the root respects $G=(V,E)$4, while the root’s centrality is set to the minimum in the graph to avoid over-preserving trivial star edges (Cui et al., 10 Aug 2025). This detail matters because the method simultaneously exempts the root node from deletion yet avoids making all root-adjacent edges effectively immutable.

The low-resource formulation in (Lin et al., 2023) uses a different augmentation regime. Its three event-level strategies are adversarial attack, feature dropout, and Graph DropEdge. Here the perturbations operate on event representations or graph connectivity for target-wise contrastive learning rather than through centrality-calibrated nodewise probabilities. The contrast underscores that “adaptive” in (Cui et al., 10 Aug 2025) is specifically tied to graph topology and centrality, not merely to the use of multiple augmented views.

4. Encoder architecture and representation learning objective

The encoder in (Cui et al., 10 Aug 2025) is a standard 2-layer GCN following Kipf and Welling:

$G=(V,E)$5

where $G=(V,E)$6. Final node embeddings are pooled into a graph representation

$G=(V,E)$7

During training, each original tree is augmented twice, for example by node-drop plus edge-drop in one view and node-drop plus attribute-mask in the other, yielding embeddings $G=(V,E)$8 and $G=(V,E)$9 (Cui et al., 10 Aug 2025).

The contrastive formulation combines supervised classification with an unsupervised graph contrastive term. Using cosine similarity,

$y\in\{N,R\}$0

the unsupervised loss is defined to pull the two views of the same graph together and push views of different graphs apart, and the total objective is

$y\in\{N,R\}$1

where $y\in\{N,R\}$2 is cross-entropy over graph labels and $y\in\{N,R\}$3 balances supervised and contrastive terms. The description explicitly states that no special projection head or extra temperature parameter is needed beyond the cosine-sim definition (Cui et al., 10 Aug 2025).

In (Lin et al., 2023), the encoder is a Multi-scale Graph Convolutional Network in which the root-claim embedding is fused back into every layer through a residual “multi-scale” mechanism. That framework imposes three losses: source and target cross-entropy, supervised contrastive loss on source events, and cross-domain contrastive alignment for target versus source. It further introduces a target-wise contrastive loss to spread few target samples uniformly on the hypersphere (Lin et al., 2023). Relative to (Cui et al., 10 Aug 2025), this places greater emphasis on source-target alignment and uniformity under low-resource transfer, whereas the 2025 RAGCL is centered on adaptive structural augmentation for standard rumor-detection benchmarks.

5. Training regime, hyperparameters, and ablations

The training configuration reported for (Cui et al., 10 Aug 2025) includes augmentation rates $y\in\{N,R\}$4 and $y\in\{N,R\}$5, both with typical values $y\in\{N,R\}$6, a contrastive weight $y\in\{N,R\}$7, Adam with learning rate approximately $y\in\{N,R\}$8, 100–200 epochs, and hidden dimension approximately 200. The two augmented views may be composed from combinations such as node-drop plus edge-drop versus node-drop plus attribute-mask. Among the supported centrality measures, PageRank is reported to often give the best speed/accuracy tradeoff, with betweenness and degree also evaluated (Cui et al., 10 Aug 2025).

The ablation record in (Cui et al., 10 Aug 2025) directly links performance differences to augmentation design, centrality choice, and graph directionality.

Ablation dimension	Reported comparison	Observation
Augmentation combos	Node-drop + edge-drop yields best; Weibo 0.962 vs random aug 0.940	Best reported augmentation pair
Centrality choice	Degree 0.953, Betweenness 0.958, PageRank 0.962	PageRank is best among listed choices
Graph direction	Top-down 0.962 vs undirected 0.955	Undirected graph linked to over-smoothing at root

These ablations clarify the method’s operative assumptions. The augmentation result supports the claim that adaptive structural perturbation is materially different from random augmentation. The centrality comparison indicates that the choice of importance measure is not incidental. The graph-direction result is especially relevant because (Lin et al., 2023) models propagation as undirected, whereas (Cui et al., 10 Aug 2025) reports that top-down yields 0.962 versus 0.955 for an undirected variant, attributing the latter to over-smoothing at the root. This suggests a methodological divergence rather than a mere implementation detail.

6. Empirical performance and benchmark behavior

The evaluation in (Cui et al., 10 Aug 2025) covers four benchmark datasets: Weibo and DRWeibo for Chinese binary rumor detection, and Twitter15 and Twitter16 for English 4-way classification with labels Non, False, True, and Unverified. Baselines include PLAN, BiGCN, UDGCN, GACL, and DDGCN. Metrics are Accuracy, Precision, Recall, and F1 for the binary setting, and Accuracy plus per-class F1 for the 4-way setting.

The main reported results are the following accuracy values:

Dataset	Best baseline	RAGCL
Weibo	0.948 (DDGCN)	0.962
DRWeibo	0.878 (DDGCN)	0.894
Twitter15	0.846 (GACL)	0.867
Twitter16	0.891 (GACL)	0.905

These results are summarized in (Cui et al., 10 Aug 2025) as evidence that RAGCL outperforms state-of-the-art methods on four benchmark rumor datasets. The paper’s conclusion attributes this to adaptive, centrality-guided graph augmentations that focus representation learning on rare but highly informative deep substructures while preserving the root.

The framework described in (Lin et al., 2023) reports a different experimental regime. Its source datasets are TWITTER and WEIBO, while the target datasets are English-COVID19, Chinese-COVID19, Cantonese-COVID19, and Arabic-COVID19. It evaluates accuracy, macro-averaged $y\in\{N,R\}$9, and per-class $y\in\{N,F,T,U\}$0, and reports gains over CNN, RNN, RvNN, PLAN, BiGCN, domain-adversarial variants, and ACLR-based methods. Reported macro-$y\in\{N,F,T,U\}$1 improvements include Chinese-COVID19 from 0.867 to 0.883, English-COVID19 from 0.687 to 0.692, Cantonese-COVID19 from 0.504 to 0.705, and Arabic-COVID19 from 0.487 to 0.687, together with the claim that the method reaches more than 90% of final performance using only the first 50 replies or approximately 4 hours of elapsed time (Lin et al., 2023). These figures situate the acronym RAGCL, within the provided material, across both standard supervised rumor benchmarks and low-resource transfer settings.

7. Relation to surrounding rumor-detection research

Within the benchmark landscape described in (Cui et al., 10 Aug 2025), RAGCL is positioned against PLAN, BiGCN, UDGCN, GACL, and DDGCN. The comparison is informative because these baselines cover Transformer-based user interaction modeling, bi-directional or undirected GCN variants with root enhancement, supervised graph adversarial contrastive learning, and dual-dynamic GCN formulations. The 2025 RAGCL therefore belongs to a line of work that treats rumor detection as structured graph classification but differentiates itself by making augmentation itself structure-aware.

The formulation associated with (Lin et al., 2023) extends the contrastive perspective in another direction. It combines propagation structure with a unified contrastive transfer framework, uses XLM-RoBERTa for shared cross-lingual post encoding, and introduces both domain-adaptive contrastive learning and target-wise contrastive learning. In that setting, the model “explicitly breaks the barriers of the domain and/or language issues” via language alignment and a domain-adaptive contrastive mechanism (Lin et al., 2023). A plausible implication is that the RAGCL label, as represented in the provided record, spans at least two research emphases: adaptive augmentation for rumor propagation trees and contrastive transfer for low-resource, cross-lingual rumor detection.

A potential misconception is that rumor propagation modeling in this area uniformly assumes deep conversation trees. The explicit contribution of (Cui et al., 10 Aug 2025) is to reject that assumption through statistical analysis showing that most nodes are shallow 1-level replies. Another potential misconception is that any contrastive augmentation on rumor graphs is equivalent to RAGCL. The provided literature indicates otherwise: (Cui et al., 10 Aug 2025) centers on node-centrality-guided adaptive perturbation in top-down RPTs, whereas (Lin et al., 2023) centers on source-target alignment and target-event uniformity under low-resource transfer. The shared contrastive paradigm does not erase these architectural and problem-setting differences.

In summary, RAGCL, in its most specific sense, is the adaptive graph contrastive learning method introduced for rumor detection in (Cui et al., 10 Aug 2025), characterized by centrality-guided node and edge perturbations, root preservation, a 2-layer GCN encoder, and a joint supervised and graph-contrastive objective. In the broader usage reflected by the provided material, the label also encompasses a contrastive transfer formulation for low-resource rumor detection (Lin et al., 2023). Both formulations treat propagation structure as a primary signal, but they operationalize that signal through different graph constructions, augmentation strategies, and optimization objectives.