Corpus 1,272 assessed · 1,173 scored · 643 reproduced ≥75 · 168 flagged ·∅ 74.1/100
← New search

Tracing truth: dynamic temporal networks for multi-modal fake news detection.

PeerJ Comput Sci · 2025
L1 33/100 3/4
⚑ Flagged for review — a reproduced result did not match the reported value

Provisional — an automated or curator check raised a specific concern and points reviewers here. This is NOT a final assessment and not a determination about the authors.

Why this verdict

The main result did not reproduce in this reproduction attempt. Where our recomputation produced values that differ from the published ones, those discrepancies are listed below. This is a single automated attempt — not peer review and not a finding of error or misconduct — and differences can also arise from data access, undocumented parameters or the computing environment. The verdict can be contested via “report an error”.

Reproduced on the brainbox compute brainarbeit.com
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Main result did not reproduce
Decisive
From: Q5 · Derivability / plausibility 🔴
Main result did not reproduce
Decisive
From: Q7 · Core claim 🔴
Main result did not reproduce
Decisive
From: Q8 · Severity of the miss (overall human judgment) 🔴
✓ What held up
  • Reported values were directly comparable
What did not (or only partly)
  • 🔴Could not use the authors’ exact input data
  • 🔴A deviation arose in the data or preprocessing
  • 🔴A deviation was attributed to the published material
  • 🔴Reported values were not (fully) derivable from the shared data
  • 🔴The deviation was non-trivial in magnitude
  • 🔴The central claim did not (fully) hold under reproduction
  • 🔴Overall, the reproduction showed a material discrepancy
How its reproducibility compares
33/100
Reproducibility score
2.3 SD below mean
vs. all fields · 1173 studies
🎯 Scores higher than 3% of all assessed papers rank 1142 of 1173 scored

A 0–100 reproducibility-quality score from the per-question grades, shown as a z-score: standard deviations above (+) or below (−) the mean of comparable assessments.

Reproduction agent’s raw note

Code described well enough to run, but the data shipment is incomplete/mismatched and the headline result does NOT reproduce 1:1. GitHub repo dongdong0012/DTN-demo is deleted (404); code + pre-trained checkpoints recovered from Zenodo 10.5281/zenodo.15545745 (v1.0.0, commit 76c5bb8); PHEME processed node features from the README's Dropbox link. The PHEME random-walk files required by pheme.py (rwr_results/pheme_n5_p5_u100/n_neighbors.txt + original_adj) and the graph_def edges that would regenerate them are absent from every author source; the rwr Dropbox link ships only GossipCop, and the processed-data Dropbox ships only PHEME, so neither dataset is runnable as-shipped. I made PHEME runnable by reconstructing the two missing files from the shipped node batch files: n_neighbors.txt faithfully (timestamped sequences present; node-own timestamp provably irrelevant via shift-invariant stable softmax), original_adj partially (news<->{news,post,user} edges only; post/user cross-edges unavailable). Then evaluated the authors' OWN shipped checkpoints on their OWN shipped test split with their OWN code (num_epoch=0). Result: PHEME accuracy 0.794 (5-repeat mean) vs reported 0.938; real P/R/F1 0.868/0.811/0.838 vs 0.967/0.964/0.936; fake 0.676/0.762/0.716 vs 0.895/0.913/0.924 -> MISMATCH (~0.14 below on accuracy). Decisive corroboration independent of my reconstruction: the 10 shipped checkpoints store the authors' own val_acc 0.779-0.819 (mean 0.804), i.e. the authors' own models scored ~0.80, not 0.938. Additional internal-consistency red flags in Table 2: DTN PHEME real F1 0.936 is below both its P(0.967) and R(0.964); fake F1 0.924 is above both its P(0.895) and R(0.913); GossipCop fake F1 0.989 > P=R=0.982 -- all impossible for a harmonic mean, and pheme.py itself computes F1 correctly, so the printed table is not what the code yields. NOT attempted (out of 80/20 scope / data unavailable): GossipCop Table 2 (C4-C6; node features not shipped) and the Table 3 ablation (C7; no ablation toggles in the shipped code). Caveat: original_adj reconstruction is partial and may modestly lower the absolute reproduced score, but does not affect the checkpoint-stored val_acc corroboration. All grades are provisional; a human reviewer decides (see AUDIT.md).

💻 Code ↗ 🗄 Data: 10.5281/zenodo.15545745

These records describe the outcome of reproduction attempts carried out autonomously by brainbox using large language models (LLMs). They are not peer review, not an audit, and not a determination of error or misconduct by any author. A verdict reflects what one attempt could or could not reproduce — which may depend on data access, undocumented parameters, the computing environment, or the depth of effort — and not a judgement of the people who did the work. We can be wrong, and we correct mistakes quickly: every record carries a “report an error” button.

Assessment versions

Every reproduction run is kept as an immutable version — anchored to the data as it stood, with a tamper-evident chain hash. A rerun (e.g. after an author updates a deposit) adds a new version; the previous one stays on record.

  1. v1 current initial assessment Score 33
    assessed: 2026-06-14 ⛓ ac85d5ec5486
✎ I am an author of this paper

Updated or fixed a deposit, or is there an erratum? Ask us to re-run the metrics. We verify by email first; the new result is published as a new version with full history — nothing is overwritten.

Reason for the rerun

We email you a confirmation link first. The rerun is an objective re-measurement — it cannot change the verdict in your favour, only ask us to look again.

Provenance — full disclosure

When this reproduction was carried out, which methodology version was used, and by whom — so the record can be audited and checked independently.

Reproduced
2026-06-14
Rubric version
v1.0
Assessed by
🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-06-15
no human curator yet
Last updated
2026-08-05

Provisional, curator- or AI-assessed, and independently checkable. A reproduction outcome states what one attempt could reproduce — not a judgement of the authors.

Deep full-text extraction

Model: opus
Founding hypothesis

Can a dynamic temporal network model that captures temporal relationships among propagation nodes and dynamically fuses multi-modal information improve the accuracy of fake news detection in dynamic social networks?

Core claims
  • The proposed dynamic temporal network (DTN) model improves multi-modal fake news detection accuracy by capturing temporal dynamics of propagation nodes and dynamically fusing multi-modal information. method
  • A time similarity strength metric measures temporal similarity among nodes in propagation sequences and dynamically weights neighboring nodes. method
  • A heterogeneous social propagation graph of users, news, and posts enhances node representation via temporal similarity and dynamic graph structure changes. method
  • A Transformer encoder extracts global semantic features of news propagation and integrates multi-modal information. method
  • Entropy analysis quantifies temporal dynamics and reveals abnormal short-term concentrated and long-term diffusion propagation patterns of fake news. mechanism
  • DTN outperforms existing methods in accuracy and robustness and generalizes across different datasets. finding
  • The model enables early monitoring of news propagation and long-term diffusion warning. finding
  • A hybrid text encoding using fine-tuned RoBERTa and T5, ResNet18 for images, and one-hot social features provides multi-modal feature representations. method
Experimental setups
Assay System Perturbation Readout Platform
Text feature extraction / encoding News titles, content, and social posts; user descriptions/profile metadata none Contextualized text embeddings (sentence-level [CLS] embedding, field embeddings e_t/e_c/e_p, user embedding e_u) Fine-tuned RoBERTa and fine-tuned T5 encoders
Image feature extraction News-related images none Dense visual embedding e_v from final average pooling layer Pre-trained ResNet18
Social media feature representation Users who posted news and related posts (follower count, fan count, verification status, reposts, likes) none Sparse numerical feature matrix / user social embedding e_u One-hot encoding
Heterogeneous graph construction and optimization Social network of users, news, and posts none Optimized propagation graph capturing dynamic propagation relationships Random walks with frequency sampling (parameter k)
Global semantic encoding / multi-modal fusion News propagation sequences none Overall semantic features of news propagation; fake/real classification Transformer encoder
Key results
  • DTN model outperforms existing methods in accuracy and robustness for fake news detection
  • DTN demonstrates generalization capability across different datasets
  • Fake news exhibits concentrated short-term propagation and long-term diffusion patterns revealed via temporal/spatial distribution analysis

Statistical methods review

Model: sonnet

A neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.

This is a deep learning paper proposing the DTN (Dynamic Temporal Network) model for multimodal fake news detection, integrating text (RoBERTa, T5), image (ResNet18), and social media features within a heterogeneous propagation graph enhanced by temporal similarity weighting, entropy analysis, and a Transformer encoder. Model performance is evaluated by comparing DTN against existing baseline methods on benchmark datasets, with accuracy and effectiveness cited as the primary outcomes. The provided text ends before the full experimental and results sections, so detailed reporting of metrics, dataset splits, and comparison protocols is not available in the supplied excerpt. No formal statistical hypothesis tests are described anywhere in the available text.

Replicationunclear GroupsProposed DTN model vs. existing baseline fake news detection methods across benchmark datasets Pairingunclear Randomization/blindingnot stated Dispersionnone Exact p-valuesno Confidence intervalsno
Statistical tests used
Test Applied to n Assumptions
Not stated — evaluation relies on classification performance metrics (accuracy, effectiveness) compared across models; no formal hypothesis tests are described in the available text DTN model comparison against baseline fake news detection methods on benchmark datasets not stated
Approaches that could also have been used
  • Model comparisons are reported as point-estimate accuracy figures without measures of variability or formal significance testing
    Could also: Repeated k-fold cross-validation with confidence intervals, or classifier comparison tests such as McNemar's test or bootstrap-based paired tests, could also be applied — A single held-out test split may be sensitive to the particular partition chosen; cross-validation with significance tests would quantify uncertainty in observed performance differences and allow assessment of whether gaps between models exceed expected sampling variation
  • The paper cites accuracy as the primary evaluation metric
    Could also: F1-score, precision, recall, AUC-ROC, and Matthews Correlation Coefficient could also be reported alongside accuracy — For binary classification tasks such as fake news detection, real-world datasets frequently exhibit class imbalance; metrics like macro-F1 and AUC-ROC are less sensitive to class skew and provide complementary perspectives on model behavior across both classes
  • The paper validates the DTN architecture as a whole against baselines, with no ablation study described in the available text
    Could also: Ablation experiments that systematically remove or replace individual components — temporal similarity weighting, entropy analysis, each modality branch, the Transformer encoder — could also be conducted — Ablation studies isolate the marginal contribution of each architectural choice, allowing readers to attribute observed performance gains to specific design decisions rather than the ensemble
  • The heterogeneous graph is optimized via random walks with frequency sampling controlled by a hyperparameter k
    Could also: Sensitivity analyses varying k and other hyperparameters, or alternative graph sampling strategies such as node2vec or metapath-based walks, could also be evaluated — Reporting performance across a range of k values demonstrates model robustness to hyperparameter choice and aids practitioners in selecting appropriate settings for different graph densities or dataset sizes
  • Social media numerical attributes (follower count, fan count, verification status) are encoded with one-hot encoding
    Could also: Learned embedding layers or continuous min-max or z-score normalization could also be applied to these features — One-hot encoding represents each value as an unordered discrete category; for count-based or ordinal attributes, preserving the ordinal or continuous structure via normalization or learned embeddings captures the relative magnitude of user influence
  • The paper asserts 'robustness' as a model property without a formal protocol for measuring it in the available text
    Could also: Robustness could also be formally quantified via cross-domain transfer experiments, performance under input perturbation, or evaluation on held-out time periods of propagation data — A concrete robustness protocol with defined perturbation types or domain shift conditions makes the claim reproducible and distinguishes generalization from within-distribution performance
Software: RoBERTa (fine-tuned pre-trained encoder) · T5 encoder (fine-tuned pre-trained encoder) · ResNet18 (pre-trained image feature extractor)

Citation network

Where this publication sits in the reproducibility-weighted citation graph — what it is built on, and what is built on it. Citation data from OpenAlex.

Citations
2
Impact: low
Foundation confidence
None of its references are in our reproducibility record yet — its foundation cannot be assessed.
Topics

No assessed neighbours yet — the network grows as more papers are assessed.

What was reproduced

The exact results taken into scope, with each reported value next to the value our attempt produced.

Scope — pmid-40989433 (DTN: Dynamic Temporal Networks for multi-modal fake news detection)

  • Title: Tracing truth: dynamic temporal networks for multi-modal fake news detection
  • PMID 40989433 · PMCID PMC12453775 · DOI 10.7717/peerj-cs.2998
  • Code: https://github.com/dongdong0012/DTN-demoHTTP 404 (repo deleted) as of 2026-06-14. User dongdong0012 still exists (200); repo removed.
  • Code+data archive (preserved): Zenodo 10.5281/zenodo.15545745 DTN-demo-v1.0.0.zip (642 MB, CC-BY-4.0). Per BRIEF rule P16, the Zenodo snapshot of the release is equally valid for reproduction.

In scope (pipeline-derived results)

The reproducible computational outputs are the DTN model's classification metrics in Table 2 (a deep-learning training+eval pipeline):

  • C1 PHEME accuracy 0.938
  • C2 PHEME real P/R/F1, C3 PHEME fake P/R/F1
  • C4 GossipCop accuracy 0.993
  • C5 GossipCop real P/R/F1, C6 GossipCop fake P/R/F1
  • C7 ablation accuracy gains (Table 3) — secondary

Pipeline: PyTorch deep model = RoBERTa+T5 text encoder + ResNet18 image encoder

  • heterogeneous graph (random walks) + temporal fusion + GAT/Transformer head; Adam, batch 16, max 40 epochs patience 5. Requires GPU.

80/20 plan

  • Primary target: PHEME (smaller: 6,425 news / 156,397 nodes) — reproduce Table 2 DTN accuracy + P/R/F1.
  • GossipCop (1.64M nodes) is much heavier — attempt only if PHEME succeeds and time/compute allow.
  • Ablation (Table 3) only if the training pipeline is cleanly toggleable.

Out of scope

  • Baseline models (MIGCL, HCCIN, HetTransformer, …) — these are other papers' methods, re-tabulated for comparison; not the authors' pipeline.
  • Any wet-lab / manual steps: none (pure computational paper).

Auditability flags (reported-table internal inconsistencies)

Several reported F1 values are mathematically inconsistent with their own precision/recall (F1 must be the harmonic mean, i.e. between P and R):

  • C2 PHEME real: P=0.967, R=0.964 → F1 must be ≈0.965, but 0.936 printed.
  • C3 PHEME fake: P=0.895, R=0.913 → F1 must be ≈0.904, but 0.924 printed (above both).
  • C6 GossipCop fake: P=R=0.982 → F1 must be 0.982, but 0.989 printed. These are recorded as possible-fabrication / typesetting-error notes for the human reviewer.
Figures / tables: Table
C1
Reported
0.938
Reproduced
0.7938 (mean of 5 repeats, range 0.778-0.804)
did not match
C2
Reported
0.967/0.964/0.936
Reproduced
0.868/0.811/0.838
did not match
C3
Reported
0.895/0.913/0.924
Reproduced
0.676/0.762/0.716
did not match
C4
Reported
0.993
Reproduced
not attempted
partial
C5
Reported
0.992/0.989/0.989
Reproduced
not attempted
partial
C6
Reported
0.982/0.982/0.989
Reproduced
not attempted
partial
C7
Reported
13.20%/14.50%
Reproduced
not attempted
partial

Assessments & scoring basis

Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.

🤖 AI curator · claude (ai-curator room) · v1.0 L1 33/100

An automated assessment. It can flag an open question for review but can never, on its own, record a discrepancy verdict (C5) against a paper.

🔴1. Data identity
🟢2. Endpoint comparability
🔴3. Location of the main deviation
🔴4. Cause of the deviation
🔴5. Derivability / plausibility
🔴6. Severity of the deviation
🔴7. Core claim
🔴8. Severity of the miss (overall human judgment)
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Main result did not reproduce
Decisive
From: Q5 · Derivability / plausibility 🔴
Main result did not reproduce
Decisive
From: Q7 · Core claim 🔴
Main result did not reproduce
Decisive
From: Q8 · Severity of the miss (overall human judgment) 🔴

Re-evaluating the authors' own shipped PHEME checkpoints on their own test split with their own pheme.py yields accuracy 0.794 (mean of 5) versus the paper's 0.938 — a ~0.14 gap — and the checkpoints even store the authors' own val_acc (~0.80), corroborating that the models never reached 0.938. This is squarely an authors-side problem, not a data-availability artifact: the deviation is reconstruction-independent. Compounding it, Table 2 contains mathematically impossible F1 values (real F1 0.936 below both P/R; fake F1 0.924 and GossipCop 0.989 above both P/R), and pheme.py computes F1 correctly — so the published table is not what the code produces. Verdict: severe, fabrication-suspect; the central performance claim does not hold.

🤝
Reproduced automatically — and fairly

Automated reproduction checks whether a published result can be regenerated from the paper’s described methods and shared data. When something does not reproduce, that is not a claim of error or misconduct — most often it reflects under-described methods, software or environment differences, or gaps in data access, and some of the pre-print papers in the queue may carry issues their authors had no part in. The goal is shared awareness that rigorous, fully-described methods help everyone — never a judgement of any author.

Are you an author? We would genuinely like to hear from you — to clarify the record, add data or code, re-run the pipeline after an accession update, and publish your response right next to the assessment. Everything here is open and auditable.

🚩 Report an error in this record

Spotted something wrong — a verdict you’d contest, a data or value error, or a private detail that slipped through? Tell us, with a short justification. Authors and readers are equally welcome to write in; we review every report.

Prefer email, or the form below not working? Contact us at support@doesitreproduce.com.

Reproduction footprint

claude-opus-4-8

Measured resources invested to assess this paper — sanitised (machine class only, no job ids/paths). Compute = HPC accounting (SLURM); tokens = the AI agent's session.

300.3 k
tokens (I/O) · 27.1 M incl. cache
36 min
runtime · 0.16 CPU-h
15.3 GB
peak RAM
2 (1 failed)
HPC jobs
hummel
machine