Graph Neural Networks for Misinformation Detection: Performance-Efficiency Trade-offs
Authors: Soveatin Kuntur, Maciej Krzywda, Anna Wróblewska, Marcin Paprzycki, Maria Ganzha, Szymon Łukasik, Amir H. Gandomi
Organizations: Warsaw University of Technology, plac Politechniki 1, 00-661 Warsaw, Poland · Corresponding author · AGH University of Krakow, Faculty of Physics and Applied Computer Science, Krakow, Poland · Systems Research Institute, Polish Academy of Sciences, Newelska 6, 01-447 Warsaw, Poland · Faculty of Engineering and IT, University of Technology Sydney, 5 Broadway, Ultimo NSW 2007, Australia · University Research and Innovation Center (EKIK), Óbuda University, Bécsi út 96/B, Budapest, 1034, Hungary · Department of Computer Science, Khazar University, Baku, Azerbaijan
The rapid spread of online misinformation has led to increasingly complex detection models, including large language models and hybrid architectures. However, their computational cost and deployment limitations raise concerns about practical applicability. In this work, we benchmark graph neural networks (GNNs) against non-graph-based machine learning methods under controlled and comparable conditions. We evaluate lightweight GNN architectures (GCN, GraphSAGE, GAT, ChebNet) against Logistic Regression, Support Vector Machines, and Multilayer Perceptrons across seven public datasets in English, Indonesian, and Polish. All models use identical TF-IDF features to isolate the impact of relational structure. Performance is measured using F1 score, with inference time reported to assess efficiency. GNNs consistently outperform non-graph baselines across all datasets. For example, GraphSAGE achieves 96.8% F1 on Kaggle and 91.9% on WELFake, compared to 73.2% and 66.8% for MLP, respectively. On COVID-19, GraphSAGE reaches 90.5% F1 vs. 74.9%, while ChebNet attains 79.1% vs. 66.4% on FakeNewsNet. These gains are achieved with comparable or lower inference times. Overall, the results show that classic GNNs remain effective and efficient, challenging the need for increasingly complex architectures in misinformation detection.
Figures & tables
Dataset
Lang.
Size
Balance
Labels
Curation
CLICK-ID [ 28 ]
ID
15K
I
Clickbait/Non-clickbait
H
LIAR [ 27 ]
EN
12.8K
I
Fine-grained truth
H
FakeNewsNet [ 21 ]
EN
23.9K
I
Fake/Real
A
Kaggle ISOT [ 1 ]
EN
44.9K
B
Fake/Real
A
WELFake [ 26 ]
EN
72K
B
Fake/Real
A
COVID-19 [ 22 ]
EN
10K
I
Fake/Real
A
Table 1: Summary of datasets used. Lang. : language. Balance : B = balanced, I = imbalanced. Curation : H = human-annotated, A = automatic.
Dataset
Model
Precision (%)
Recall (%)
F1 (%)
AUC-ROC (%)
MCC (%)
Inference (ms)
CLICK-ID
LogReg
50.3 ± 21.3
50.3 ± 21.4
50.3 ± 21.3
51.5 ± 21.6
13.1 ± 5.3
5.85
SVM
50.2 ± 21.4
51.3 ± 22.1
49.4 ± 21.1
50.9 ± 21.8
11.9 ± 4.6
6.95
MLP
58.8 ± 25.6
58.5 ± 25.6
57.0 ± 24.9
61.9 ± 26.9
28.5 ± 12.3
10.84
GAT
70.1 ± 1.8
69.6 ± 2.1
69.5 ± 1.9
75.2 ± 1.2
38.3 ± 3.6
5.24
COVID-19
LogReg
72.1 ± 31.3
72.0 ± 31.3
71.9 ± 31.2
77.8 ± 33.9
58.3 ± 24.8
0.92
SVM
72.1 ± 31.5
71.9 ± 31.4
71.9 ± 31.4
78.2 ± 34.1
58.2 ± 25.1
2.70
Table 2: Performance comparison across datasets. Results are reported as mean ± standard deviation over multiple runs. F1 score is the primary metric. MCC is reported to account for class imbalance. LogReg denotes Logistic Regression, MLP denotes Multilayer Perceptron, and SVM denotes Support Vector Machine.
Dataset
Best GNN
Precision (%)
Recall (%)
F1 (%)
Inference (ms)
Kaggle
GraphSAGE
96.8 ± 0.2
96.8 ± 0.2
96.8 ± 0.2
11.11
WELFake
GraphSAGE
91.9 ± 0.2
91.9 ± 0.2
91.9 ± 0.2
16.21
COVID-19
GraphSAGE
90.5 ± 0.8
90.5 ± 0.8
90.5 ± 0.8
7.18
MPID
ChebNet
88.9 ± 0.5
88.9 ± 0.5
88.6 ± 0.5
20.68
FakeNewsNet
ChebNet
79.3 ± 2.1
80.5 ± 1.8
79.1 ± 1.7
8.40
CLICK-ID
GAT
70.1 ± 1.8
69.6 ± 2.1
69.5 ± 1.9
5.24
Table 3: Best-performing classic GNN per dataset, selected by F1 score after hyperparameter optimization. Results are reported as mean ± standard deviation.
Dataset
F1@10%
F1@20%
F1@30%
Drop (%)
CLICK-ID
65.0
69.0
69.5
6.5
COVID-19
87.2
90.3
90.5
3.6
FakeNewsNet
74.2
79.1
79.1
6.2
Kaggle
93.5
95.7
96.8
3.4
LIAR
59.2
57.3
58.8
0.7
MPID
85.6
88.6
88.4
3.2
Table 4: Effect of training data size on the performance of classic GNNs (F1 score).
Figure 1: Performance–efficiency trade-off across learning algorithms. Each point corresponds to the best-performing configuration of a model on a given dataset. The x-axis reports inference time (ms), while the y-axis shows F1 score (%). All GNN variants are shown using the same color to emphasize family-level behavior, with algorithm names annotated for clarity.