acceptodds
Under review as a conference paper at ICLR 2027

CPNS: Curriculum Learning-Guided Progressive Key Node Selection for Fake News Detection

Abstract

Selecting valuable nodes from the social propagation graph is crucial for fake news detection. Previous work mainly evaluates node importance based on propagation depth, then constructs propagation subgraphs to mitigate noise interference. However, they still suffer from two limitations: (1) propagation subgraphs make signals from both valuable and noisy nodes smoother; (2) affected by depth, deep nodes are easily ignored. To address these issues, we propose Curriculum Learning-guided Progressive Key Node Selection for Fake News Detection (CPNS) to select key propagation nodes. Specifically, we propose a curriculum learning-guided progressive propagation strategy to select deep nodes. Moreover, we weight nodes based on properties, thereby distilling informative knowledge from selected nodes. Finally, we introduce multi-frequency consistency learning. It jointly models low-frequency propagation consensus and high-frequency local differences, thereby avoiding excessive smoothing of both noisy and valuable nodes. Experimental results demonstrate that our approach achieves state-of-the-art performance on two benchmark datasets, improving by 1.54% and 2.15%, respectively.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Fake News Detection, GNNs, Curriculum Learning, Multi-frequency Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_g8wsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CPNS for propagation-based fake news detection, aiming to make better use of informative deep nodes and reduce over-smoothing during aggregation. CPNS selects candidate nodes based on propagation morphology and weights them using LLM evidence scores. It progressively exposes propagation edges through a depth curriculum, uses multi-branch GraphSAGE routing, and applies consistency constraints across the main, low-frequency, and high-frequency residual views. The main graph representation is combined with the source-news embedding and morphology features to predict news veracity.

Critical strengths
  1. The research problem is relevant. The paper examines the use of deep propagation nodes and over-smoothing during aggregation, both of which may limit how effectively deeper propagation information contributes to fake news detection.
  2. The framework design is reasonable. Gradually exposing propagation edges introduces deeper information during training, while incorporating local consensus and node differences allows the model to consider shared information and variation among nodes.
  3. Multiple aspects are covered in the experiments. Comparisons include multiple baselines on two public datasets. Analyses of candidate selection, curriculum learning, routing signals, and robustness to feature noise help assess component contributions and performance under noisy conditions.
Critical weaknesses
  1. The methodological novelty is limited. The method mainly combines candidate node selection, LLM weighting, curriculum learning, multi-branch routing, and multi-frequency constraints, without clearly explaining the specific improvements over existing methods. The paper could explain more clearly how the modules work together and provide further evidence for the benefits of combining them.
  2. The evidence supporting the proposed mechanisms is not sufficient. Results do not directly show reduced reliance on depth, better use of deep-node evidence, or less over-smoothing. Further analysis is needed to assess deep-node contributions, the relationship between depth and node value, and over-smoothing with and without the multi-frequency constraints.
  3. LLM scoring reliability remains unvalidated. The paper does not explain how hallucinations or incorrect judgments are identified and controlled, and lacks explicit scoring criteria, comparisons with human annotations, and error analysis.
  4. The ablation studies could be more comprehensive. The paper does not clearly report ablations removing LLM weighting, isolating the contribution of adaptive multi-branch routing, or separately removing the low- and high-frequency constraints.
  5. Baseline descriptions are inconsistent, and some metric values require verification. Baselines in Section 4.2 do not fully match the result tables, including GAMC, CGAT, HCSL, SEE-GraphMAE, and Semantic-GMAE. In Table 3, UPFD-SAGE’s Macro-F1 (0.8591) exceeds its accuracy (0.8579), contradicting the formulas in Appendix B.3 for the same binary predictions. These values need checking.
Questions for authors
  1. What are the specific new mechanisms introduced beyond existing curriculum-learning and low-/high-frequency graph-learning methods?
  2. Could you provide analyses that directly support reduced reliance on depth, preservation of deep-node evidence, or mitigation of over-smoothing?
  3. What criteria does the LLM use to score node evidence value and reliability? How are hallucinations and incorrect judgments identified or controlled during scoring?
  4. Could you provide results removing LLM weighting, replacing adaptive routing with a single branch or a fixed mixture, and separately removing the low- and high-frequency constraints to isolate the contributions of these components?
  5. Please clarify the baseline discrepancies involving GAMC, CGAT, HCSL, SEE-GraphMAE, and Semantic-GMAE, and report UPFD-SAGE’s per-run Macro-F1 and accuracy in Table 3, including how they are computed and aggregated.
4 comments
AuthorsOpenReview ↗

Response to Reviewer g8ws — Critical Weaknesses(W1–W3)

W1 Response

Thank you for questioning the specific improvements over existing methods. We acknowledge that the original description did not clearly explain how the modules work together. We revised the opening of Section 3 to explain the modules’ roles and connections.

[Details]:

In the opening of Section 3.

"The three components operate at the stages of information introduction, utilization, and representation learning, respectively. The opening explains that information is introduced, weighted, and represented through the three connected components."

W2 Response

Thank you for questioning reduced reliance on depth and better use of deep-node evidence. You also ask about oversmoothing with and without multi-frequency constraints. The manuscript analyzes deep-node contributions and the relationship between depth and node value. It also analyzes oversmoothing with and without multi-frequency constraints. Appendix A adds human assessment of node value. Appendix F explains the evidence concerning oversmoothing.

(1) Analyze the relationship between depth and node value

Sampling groups cover datasets, splits, and depth levels. There are 12 groups. Each group contains 250 nodes. The total is 3,000 non-root nodes. Shallow nodes have depth one or two. Deep nodes have depth at least three. Five experts participated in assessment. A valuable node must help judge the root news. Topic overlap alone is insufficient. This process is separate from LLM scoring.

[Details]:

In Appendix A.

"A valuable node must provide information useful for judging the root news. Topic relevance alone is insufficient. The sampling covers 12 dataset-split-depth groups, with 250 nodes per group and five expert assessors."

In Appendix F.

"The cited studies do not establish prediction degradation in our two-layer CPNS. The appendix describes the observation as a potential risk and says direct experiments are still needed."

W3 Response

Thank you for questioning the LLM scores. The original manuscript inadequately explained how hallucinations or incorrect judgments are identified and controlled. Appendix B therefore provides assessment criteria, output mappings, and usage limits.

(1) Define explicit scoring criteria

Evidence must help judge the root news. Topic overlap alone is insufficient. Fact-checking keywords alone are also insufficient. Outputs distinguish evidence, noise, and ambiguous information.

(2) Explain the identification and control of hallucinations and incorrect judgments

Requests exclude ground-truth labels. Reasons are not classifier inputs. Scores only affect bounded node weights. The assessor does not produce the final classification.

[Details]:

In Appendix B.

"The assessor distinguishes evidence, noise, and ambiguous information, and returns value and reliability scores. The output includes evidence, noise, or ambiguous labels, together with value and reliability scores."

In Appendix B.

"The classifier receives neither generated veracity judgments nor rationale text. The classifier uses supervised parameters to make final veracity predictions; rationales are not classifier inputs."

AuthorsOpenReview ↗

Response to Reviewer g8ws — Critical Weaknesses(W4–W5)

W4 Response

Thank you for requesting separate ablations of LLM weighting, routing. We conducted new ablations of LLM weighting and multi-branch routing. Appendix G now adds a comparison with LLM scoring disabled.

(1) Isolate the effect of LLM weighting

Only the evidence coefficient is set to zero. The candidate-prior coefficient remains 0.25. Candidate weights become 1.25. Other node weights remain 1.

(2) Isolate the contribution of adaptive multi-branch routing

The manuscript compares training with and without the curriculum. It also compares candidate selection methods. Other comparisons examine depth and uncertainty signals. Appendix G reports the evidence-score control.

[Details]:

In Appendix G.

"The control disables LLM evidence scoring while retaining the other settings. The classifier receives no separate graph-level LLM statistics. Thus, no other input path supplies LLM assessment information when the coefficient is zero."

In Appendix D.5.

"Gains vary across external models, so the effect depends on encoder structure and data distribution. With MFC, the PolitiFact degradation at sigma=0.10 and 0.20 decreases from 20.09% and 45.85% to 5.22% and 17.68%, respectively."

W5 Response

Thank you for identifying inconsistent baseline descriptions and the UPFD-SAGE metric discrepancy in Table 3. Its reported Macro-F1 exceeds its accuracy. We checked the baseline descriptions and metric values, and reran the experiments. We revised Section 4.2 and Appendix D.4, and added HCSL results.

(1) Align baseline descriptions with the result tables

Baseline additions. We added a GAMC description. We removed the unmatched CGAT description.

(2) Clarify HCSL baseline descriptions and results

Tables 2–3 add overall results. Tables 4–5 add noise results.

[Details]:

In Section 4.2, Appendix D.4, and Tables 2--5.

"We compare CPNS with two categories of representative graph-structured fake news detection methods. Detailed descriptions of these baselines are provided in Appendix D.4. Detailed descriptions of these baselines are provided in Appendix D.4."

AuthorsOpenReview ↗

Response to Reviewer g8ws — Questions for Authors (Q1–Q3)

Q1 Response

Thank you for asking about mechanisms beyond existing curriculum and low-/high-frequency graph-learning methods. We do not present individual modules or methods as new contributions. Instead, we combine and adapt them to use deep-node evidence and preserve local differences. The original text inadequately distinguished established tools from their arrangement in our method. Section 3 now explains how the curriculum, weights, and frequency objectives connect.

(1) Explain the mechanisms beyond existing curriculum-learning methods

The depth curriculum gradually exposes propagation relations. Depth sets the learning order. It does not directly define node value. Evidence weights adjust node contributions. Multiple branches provide different aggregation paths. Routing determines their mixture.

(2) Explain the mechanisms beyond existing low-/high-frequency graph-learning methods

The low-frequency objective targets shared information. The high-frequency objective targets local differences. Both contribute during training.

[Details]:

In Section 3.

"The curriculum, evidence-aware node weighting, routing, and frequency objectives are described as connected components. The opening of Section 3 assigns the curriculum to information access, evidence weights to node contributions, and frequency learning to representation differences."

Q2 Response

Thank you for requesting direct analyses of depth reliance, deep-node evidence preservation, and oversmoothing. We used human verification to assess node value. The manuscript presents these assessment results. Appendix A describes human assessment, while Appendix F discusses oversmoothing.

(1) Analyze the preservation of deep-node evidence

Sampling groups cover datasets, splits, and depth levels. There are 12 groups. Each group contains 250 nodes. The total is 3,000 non-root nodes. Shallow nodes have depth one or two. Deep nodes have depth at least three. Five experts participated in assessment.

[Details]:

In Appendix A.

"Five experts assessed sampled nodes using whether each node helps determine the root news veracity. A valuable node must provide information useful for judging the root news; topic relevance alone is insufficient."

In Appendix F.

"The analysis describes a potential risk and does not establish preferential loss of evidence from deep valuable nodes. The appendix explicitly says this does not establish prediction degradation in the two-layer CPNS."

Q3 Response

Thank you for asking about node evidence value, reliability criteria, and control of hallucinations. We specify the evidence assessor’s inputs and outputs. Appendix B now defines criteria for node evidence value and reliability. It also specifies how these scores affect node weights. It also explains how hallucinations and incorrect judgments are identified or controlled during scoring.

(1) Specify the criteria for node evidence value and reliability

Evidence must help judge the root news. Topic overlap alone is insufficient. Fact-checking keywords alone are also insufficient. Outputs distinguish evidence, noise, and ambiguous information.

(2) Explain the identification and control of scoring errors

Requests exclude ground-truth labels. Reasons are not classifier inputs. Scores only affect bounded node weights. The assessor does not produce the final classification.

[Details]:

In Appendix B.

"Assessment uses the root news, candidate text, local context, and morphology summaries; the graph label is excluded. The assessor receives readable text and structural summaries, not only numerical feature vectors."

In Appendix B.

"Unavailable assessments map to a signed value of zero. The returned categories are evidence, noise, and ambiguous; unavailable assessments receive signed value zero."

AuthorsOpenReview ↗

Response to Reviewer g8ws — Questions for Authors (Q4–Q5)

Q4 Response

Thank you for requesting controls that separately isolate LLM weighting, routing, and frequency constraints. Appendix G now adds control settings and results with LLM evidence scoring disabled. The control retains candidate priors, node features, model structure, and other training objectives. It only disables the scores’ influence on node readout. Appendix D.5 further reports the overall effect of multi-frequency learning under noise. The low-frequency view describes shared propagation information, while high-frequency residuals describe local differences. Both receive separate label-conditioned contrastive supervision.

[Details]:

In Appendix G.

"The control disables the contribution of LLM evidence scores to node weights. Both conditions retain the same data splits, node features, candidate sets, and trainable parameter count."

In Appendix D.5.

"Gains vary across external models, so the effect depends on encoder structure and data distribution. Tables 4 and 5 report the complete per-backbone robustness results, making the cross-model pattern explicit."

Q5 Response

Thank you for identifying baseline discrepancies and requesting UPFD-SAGE’s per-run Macro-F1 and accuracy. We reran the experiments and obtained results. We acknowledge carelessness when entering the original metric values. We standardized baseline descriptions and added missing methods to the tables.

(1) Clarify baseline discrepancies involving GAMC, CGAT, HCSL, SEE-GraphMAE, and Semantic-GMAE

The manuscript adds a GAMC description. It removes the unmatched CGAT description. It standardizes the Semantic-GMAE name. Tables 2–3 include overall results. Tables 4–5 include noise results.

[Details]:

In Section 4.2, Appendix D.4, and Tables 2--5.

"We compare CPNS with two categories of representative graph-structured fake news detection methods. Detailed descriptions of these baselines are provided in Appendix D.4. Tables 3 and 4 compare CPNS with all baselines on two highly noisy datasets."

Reviewer_kh4SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CPNS for propagation based fake news detection. The method combines morphology conditioned candidate selection and LLM based node weighting with progressive depth based edge exposure, a multi-branch GraphSAGE encoder, and auxiliary consistency objectives involving neighborhood averages and residual representations. The final prediction combines the pooled graph representation, source-news features, and morphology descriptors.

The motivation is relevant, and the reported results on GossipCop and PolitiFact are promising. However, the submission does not sufficiently specify several central components, and the reported configuration is difficult to reconcile with the claimed treatment of deep propagation nodes. The experiments also leave substantial uncertainty about which components account for the gains.

Critical strengths
  1. The paper addresses a meaningful distinction between propagation depth and information value. Avoiding indiscriminate suppression of deep nodes is a worthwhile objective.

  2. The framework attempts to balance noise suppression with preservation of local differences, rather than relying solely on removing nodes or smoothing their representations.

  3. The evaluation includes multiple graph based baselines, curriculum and candidate selection analyses, backbone comparisons, and feature noise experiments. Appendix B.2 states that results are averaged over five random seeds, which is a positive aspect of the experimental design.

Critical weaknesses
  1. Several central algorithmic components remain unspecified. The candidate scoring function Φsel\Phi_sel in Appendix A is defined only through its arguments. The actual scoring rule, candidate budget, and numerical settings of the evidence-weighting parameters are not provided. The curriculum schedule similarly lacks stage durations, transition conditions, and a clear specification of the final training and inference graphs.

The multi-frequency objective is also incomplete. Eq (12) names a label conditioned L_NCE but does not specify the positive/negative construction, projection functions, or complete loss. The statement that these choices “must match the released training configuration” does not define the method. Likewise, L_reg is described as whichever minimization-form regularizer the implementation uses, without specifying its actual form.

  1. The reported curriculum configuration does not clearly support the deep-node claim. Fig. 1 treats nodes at depth at least three as deep, but Table 9 lists only Q1 = 1 and Q2 = 2. Under Eq (5), if these are the only curriculum stages, edges incident to nodes deeper than two are never exposed to message passing. Deep-node features can still influence the prediction because the full node set is retained and pooled. However, this is different from progressively learning their propagation relations. Please clarify whether additional stages or a final full-graph stage are omitted, and which edge set is used during inference. Also clarify whether the frequency-filter neighborhood uses the masked or full graph; the latter could expose deeper relations outside the stated curriculum.

  2. The central mechanism claims are not directly established. The definition of a “valuable” node in Fig. 1 is insufficiently documented. Please specify the annotation or scoring procedure, sampling scheme, sample sizes, and meaning of the error bars. If the same LLM used by CPNS also defines node value, independent validation is particularly important. Table 1 shows decreasing differences between neighboring representations, but this alone does not establish harmful oversmoothing or preferential loss of useful deep-node information. The analysis should connect representation homogenization to discrimination and prediction errors in the actual model configuration.

  3. The contribution of LLM assistance and the provenance of node features are unclear. Section 3.2 states that the evidence assessor receives root text, candidate text, and parent child context. Appendix A instead states that it only classifies nodes from numerical representations. These describe materially different pipelines. The exact model checkpoint, input serialization, decoding settings, output schema, and treatment of failed assessments should be specified.

  4. The empirical comparisons and robustness claims require stronger support. Five run means are reported, but no standard deviations or confidence intervals are given. This matters especially for PolitiFact, which has only 62 training and 31 validation graphs. The small validation set also makes architecture and hyperparameter selection potentially unstable.

Questions for authors
  1. What are the exact candidate scoring rule, candidate budget, node weight parameters, and complete L_NCE and routing regularization objectives?

  2. Does the curriculum ever expose edges beyond depth two? What graph is used for the frequency filter and at inference?

  3. Why is β\beta set to one if the reported method combines depth and uncertainty in Eq (9)?

  4. What text or numerical features does the LLM actually receive, and how are the 4096 dimensional node features constructed?

  5. How much improvement remains without LLM assistance and with matched features, parameter counts, and auxiliary supervision?

8 comments
AuthorsOpenReview ↗

Response to Reviewer kh4s — W1 (Part 1)

Thank you for your questions. We agree that the candidate scoring rule, candidate budget, and evidence-weighting parameters were unspecified. The curriculum schedule also lacked stage durations, transition conditions, and training and inference graph specifications. Appendix B now specifies the candidate scoring rule, candidate budget, and evidence-weighting parameters. The curriculum schedule and graph specifications are also detailed in Appendix B and Table 9. (1) Define the candidate scoring rule We specify four inputs: depth, out-degree, subtree size, and keyword matches. Each normalized input receives a weight of 0.25. We then add a graph-shape bonus. Scores retain six decimal places. (2) Specify the candidate budget Each graph has at most three candidates. Smaller candidate pools retain every available node. We normally sort candidates by score. Ties follow a fixed order. Balanced graphs prioritize coverage across branches. Other nodes remain in the graph. (3) Specify the evidence-weighting parameters The candidate-prior coefficient is 0.25. The evidence coefficient is 0.50. We clip weights to [0.50, 1.50]. Other nodes retain weight 1. (4) Clarify curriculum stages and training and inference graphs Epochs 1–33 expose depth one. Epochs 34–66 expose depth two. Epochs 67–100 use the full graph. Stages change at fixed epochs. Validation and inference use the full graph. [Details]: In Appendix B. "The candidate pool excludes the root. For non-balanced graphs, nodes are sorted by decreasing score. Ties preserve local node order. The offline candidate budget is k=3; if fewer candidates are available, all are retained." In Appendix D, Table 9. "Routing uncertainty weight (beta): 0.5. The regularization coefficient is lambda_reg=0.02, and the frequency loss weight is lambda_freq=0.05."

AuthorsOpenReview ↗

Response to Reviewer kh4s — W1(continued) (Part 2)

Response Thank you for identifying the incomplete multi-frequency objective and unspecified routing regularizer. We acknowledge missing positive/negative construction, projection functions, and the complete L_NCE definition. The actual form of L_reg was also unspecified. Appendix C now defines positive/negative construction, projection functions, and the complete L_NCE. It also specifies the actual form of L_reg and all coefficients. (1) Define positive and negative samples, projection functions, and the complete loss Pairs come from the current training batch. Other graphs sharing the anchor label are positives. Graphs with different labels are negatives. We use no additional projection layer. The temperature is 0.2. The denominator includes every target-view graph. It also includes the same-graph term. That term is not a positive. We first average over positives. We then average over valid anchors. Anchors without positives are skipped. The loss is zero when no valid anchors exist. The main view is compared with the low-frequency and high-frequency views. We do not add reverse-direction losses. (2) Specify the routing regularizer L_reg Equation (36) defines the routing loss. It sums p·logπ across nodes, branches, and layers. The result is averaged over nodes and layers. Gradients remain enabled. There is no additional branch-balancing loss. The total loss includes supervised, routing, and frequency losses. The latter two coefficients are 0.02 and 0.05. [Details]: In Appendix C. "Other graphs sharing the anchor label are positives. Graphs with different labels are negatives. We use no additional projection layer. The denominator includes all target views, including the same-graph term. Anchors without same-label partners are excluded from the average." In Appendix C. "The regularizer averages over all nodes and propagation layers. No gradient is detached when computing this term. This includes nodes excluded from classification pooling during the first two stages. No gradient is detached when computing this term."

AuthorsOpenReview ↗

Response to Reviewer kh4s — W2

Thank you for asking about edges beyond depth two and the final full-graph stage. Our curriculum exposes edges beyond depth two during the final full-graph stage. The frequency-filter neighborhood uses the masked graph during the first two curriculum stages. Appendix B and Table 9 now clarify the final full-graph stage. Appendix C also distinguishes filtering and pooling ranges to avoid confusion. (1) Clarify the final full-graph curriculum stage The schedule has three stages. Epochs 1–33 expose depth one. Epochs 34–66 expose depth two. Epochs 67–100 use the full graph. Deep edges are therefore exposed during the final stage. The first two stages restrict classification pooling. Deeper nodes are excluded from that pooling. The final stage removes this restriction. (2) Distinguish masked and full graphs in frequency filtering The filter uses only current curriculum edges. It does not use unopened edges. Auxiliary pooling still retains all node states. This does not expose hidden propagation relations. [Details]: In Appendix B and Appendix D, Table 9. "For T=100, the stages span epochs 1--33, 34--66, and 67--100, respectively. Stages change at fixed epochs and do not depend on validation performance or convergence thresholds." In Appendix B. "Validation and inference always use the complete graph. Validation and inference use standard routing probabilities without Gumbel sampling; the frequency auxiliary loss is not computed."

Reviewer_LUw3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores an interesting proposal: in the propagation graph of a specific piece of misinformation, the nodes that are far away from the center node may be important. Thus, this paper aims to extract useful information from these node with reducing the propogation noisy. To this end, the authors propose CPNS to solve these questions.

Critical strengths
  • This paper explores an important research question that is helpful in practice.
  • The proposed model seems to achieve performance improvements of up to 2.15%.
Critical weaknesses
  • The authors claim that 'progressive key node selection' is critical, while it seems to be the main contribution of this paper. However, in my understanding (according to Eq. 5), CPNS gradually unlocks the edges (instead of nodes). Namely, I think that CPNS is changing the edges that are used for message passing, instead of directly knowing the critical key nodes. I'm not sure whether they are the same.
  • By the way, I think the authors are mixing the existing modules in a somewhat novel way instead of proposing a novel module. Thus, I think the authors need to answer why the methods are effective beyond performance.
  • I find that the equations of this paper and the experimental settings conflict (maybe I missed something, but I think the authors should explain).
    • In Eq. 9, the authors define α+β=1\alpha + \beta = 1. However, in Table 9, the authors claim that β=1\beta = 1. It seems that α\alpha is useless. Meanwhile, the authors do not provide a parameter analysis.
    • In Table 9, the authors claim that they set Q1=1Q_1=1 and Q2=2Q_2=2. However, it is unclear how to introduce edges whose depth is greater than 2 and when to introduce them. Since the authors' main claim is that deep nodes may be important, I think this setting is critical.
  • The authors employ LLMs to generate evidence scores and reliability scores. I think this is a vanilla method as an LLM-as-enhancer. Since the authors introduce it to improve performance, the authors should answer which parts really contribute to the performance gain.
  • The PolitiFact dataset only provides 62 instances to train. I think there might be randomness. I think the authors should provide a statistically significant test on it.
  • The authors consider HCSL as a baseline in Table 1 but do not provide its performance in Tables 4 and 5.
Questions for authors
  • What is the precise definition of node selection in CPNS, and how is it different from progressively masking propagation edges?
  • Please explain the experimental settings. If possible, the authors could provide more hyperparameter analysis.
  • If the authors could provide a more detailed ablation study or explain which parts contribute the most, the contributions of CPNS could be clearer.
  • Please refer to the weaknesses.
5 comments
AuthorsOpenReview ↗

Response to Reviewer LUw3 — Critical Weaknesses(W1–W2)

W1 Thank you for asking whether edges or nodes are unlocked. Our original description did not make this clear. Equation (5) directly controls when propagation edges become available. We clarified that it works with a depth mask for classification nodes. Together, they gradually allow the corresponding nodes to participate in classification. Node importance is further determined through evidence weighting and classification learning. Progressive selection therefore means gradually expanding node participation. Opening an edge does not automatically identify a key node. Section 3.3 now clarifies curriculum control over propagation and classification participation. Appendix B further explains that unselected nodes are retained.

(1) Distinguish progressive edge unlocking from node selection Equation (5) controls propagation edges. It determines when information becomes available. Classification pooling also expands across stages. This operation does not provide true node-value labels. The candidate set limits assessment requests. Unselected nodes are not deleted. The final stage and inference use the full graph. (2) Clarify the relationship between edge unlocking and key-node identification Evidence weights also affect node contributions. Learned node representations also matter. Edge masking alone does not identify valuable nodes.

[Details]: In Section 3.3. "The depth curriculum progressively exposes propagation edges and allows their nodes to enter classification. The hidden states retain the complete node set, so deeper nodes are not deleted or labelled as noise." In Appendix B. "The final training stage, validation, and inference use the full graph. Full-graph validation and inference use every node and do not compute the frequency auxiliary loss."

W2: Thank you for asking why the combination of existing modules is effective beyond performance. We rewrote the opening of Section 3 to explain the modules’ roles and connections. It connects information access, contribution weighting, and representation learning.

(1) Explain effectiveness beyond reported performance The depth curriculum gradually exposes propagation relations. Depth sets the learning order. It does not directly define node value. Evidence weights adjust node contributions. Multiple branches provide different aggregation paths. Routing determines their mixture. The low-frequency objective targets shared information. The high-frequency objective targets local differences. Both contribute during training. Section 4.4 provides several comparisons.

[Details]: In Section 3 and Section 4.4. "The depth curriculum gradually exposes propagation relations. Depth sets the learning order; it does not directly define node value. Graph difficulty combines normalized structural depth and routing uncertainty to determine each graph's supervised-loss weight." In Section 4.4. "Their combination better controls node propagation, showing the advantage of our adaptive routing strategy. Depth provides stable structural information, while routing uncertainty reflects the current learning state."

AuthorsOpenReview ↗

Response to Reviewer LUw3— Critical Weaknesses(W3–W4)

W3a

Response

Thank you for noting the apparent conflict between β = 1 and combining depth with uncertainty. On PolitiFact, β = 1 performs better because its propagation graphs are relatively deep and structurally diverse. This setting is dataset-specific, not a reporting error. For most datasets, β = 0.5 performs better. The appendix also reports PolitiFact results with β = 0.5.

[Details]:

In Appendix D, Table 9.

"Routing uncertainty weight: 0.5. The manuscript defines alpha+beta=1 and reports beta=0.5 in Table 9."

W3b Response

Thank you for asking how and when edges beyond depth two are introduced. Epochs 67–100 introduce these edges through the final full-graph stage. Appendix B now lists the three-stage curriculum and inference rules. Appendix C specifies that filtering uses only current curriculum edges.

(1) Explain how and when deeper edges are introduced

The schedule has three stages. Epochs 1–33 expose depth one. Epochs 34–66 expose depth two. Epochs 67–100 use the full graph. Deep edges are therefore exposed during the final stage. The first two stages restrict classification pooling. Deeper nodes are excluded from that pooling. The final stage removes this restriction. The filter uses only current curriculum edges. It does not use unopened edges. Auxiliary pooling still retains all node states. This does not expose hidden propagation relations. Both use the full graph. Routing uses ordinary probabilities without random sampling. Frequency losses are not computed.

[Details]:

In Appendix B.

"For T=100, the stages span epochs 1--33, 34--66, and 67--100, respectively. Transitions do not depend on validation performance or convergence thresholds."

In Appendix C.

"Frequency filtering uses neighbors from the current curriculum edge set. The encoder and filter use the same current edge set; auxiliary pooling still uses all node states."

W4 Response

Thank you for asking which parts contribute to the performance gain from LLM assistance. Appendix G now isolates the effect of LLM evidence and reliability scores. The control disables their contribution to node weights.

(1) Identify the components contributing to performance gains

Only the evidence coefficient is set to zero. The candidate-prior coefficient remains 0.25. Candidate weights become 1.25. Other node weights remain 1. The control is designed to retain the same node features. Candidates, parameter counts, and auxiliary losses remain unchanged. No additional graph-level scoring features are added. Table 13 reports results with evidence scoring disabled. GossipCop has a Macro-F1 of 0.9698. The value for PolitiFact is 0.8924. This control disables evidence scoring only. The text embedding model remains.

[Details]:

In Appendix G.

"Only the semantic evidence coefficient changes. The candidate-prior coefficient remains 0.25, with weight bounds of 0.50 and 1.50."

In Section 4.4.

"Figures 8 and 10 compare routing signals; Figures 9 and 11 compare curriculum settings. Figures 8 and 10 show that combining depth and routing uncertainty achieves the best Macro-F1 on both datasets. Depth provides stable structural information, while routing uncertainty reflects the current learning state."

AuthorsOpenReview ↗

Response to Reviewer LUw3 — Critical Weaknesses(W5–W6)

W5 Response

Thank you for raising randomness with only 62 PolitiFact training instances. PolitiFact is a public benchmark dataset. We also use random seeds to ensure reproducible experiments. Appendix D.2 documents five runs on fixed splits. The manuscript reports five runs. The seeds are 42–46. These runs use fixed data splits. Tables report their averages.

[Details]:

In Appendix D.2.

"Each experiment was run independently with seeds 42--46, and results were averaged over the fixed data splits. The fixed splits are used for all five seeds; the reported results are averages across those runs."

W6 Response

Thank you for pointing out this omission. Reporting HCSL only in some analyses left the performance comparison incomplete. We added HCSL-UPFD to Tables 2–5, covering overall and noise results.

(1) Report HCSL performance in Tables 4 and 5

Tables 2–3 now include HCSL-UPFD. They report results for both datasets. Tables 4–5 also include corresponding rows. Noise standard deviations are 0.05 and 0.10. Table 4 reports 0.855 and 0.699. Table 5 reports 0.812 and 0.720.

[Details]:

In Section 4.2 and Tables 2--5.

"We compare CPNS with two categories of representative graph-structured fake news detection methods. The listed propagation baselines include Bi-GCN, UPFD, GAMC, EBGCN, FGDGNN, and MTAS."

In Appendix D.4.

"HCSL combines intra-sample curriculum learning with hierarchical semantic learning. It captures semantic evidence at multiple representation levels and reduces the influence of less informative components."

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.