Graph Neural Networks for Jamming Source Localization

Dania Herzalla; Martin Andreoni; Willian T. Lunardi

REVIEW 3 major objections 5 minor 31 references

Graph Neural Networks for Jamming Source Localization

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Graph neural networks locate jamming sources from noisy signal maps, beating geometry-based baselines by treating localization as graph regression.

desk verdict First GNN for jammer localization with a genuinely useful confidence-fusion trick; the central claim holds within its simulated LDPL world, but not beyond it. read the letter →

arxiv 2506.03196 v2 pith:X6DS5RG3 submitted 2025-06-01 cs.NI cs.CRcs.ITcs.LGeess.SPmath.IT

Dania Herzalla , Willian T. Lunardi , Martin Andreoni This is my paper

classification cs.NIcs.CRcs.ITcs.LGeess.SPmath.IT

keywords Graph-basedlearningGraphNeuralNetworksregressionWirelesssecurityJammingsourcelocalizationAttentionmechanismConfidenceweightingRadiopropagationmodeling

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that locating a wireless jammer can be treated as an inductive graph regression problem, and that a graph neural network built for this task outperforms standard geometric and optimization-based methods. The proposed framework, CAGE, builds a graph of signal measurements, uses attention-based message passing to aggregate local and global spatial information, and blends the network's prediction with a weighted-centroid prior through a learned confidence weight. Across simulated static and dynamic radio environments, the paper reports lower localization error than all baselines, with the largest gains in sparse sampling and when the jammer lies outside the sampled region. If this holds, it would make graph neural networks a practical tool for locating jammers and other interference sources from noise-floor measurements without specialized ranging hardware.

What carries the argument

The load-bearing mechanism is CAGE (Confidence-guided Adaptive Global Estimation), built on an augmented graph $G' = (V', E')$. Each node in $V'$ is a signal measurement, edges are K-nearest-neighbor connections weighted by an exponential decay in distance, and node features carry normalized noise floor, spherical and Cartesian coordinates, and local noise statistics. A graph attention encoder produces node embeddings, and a supernode encodes the weighted centroid localization (WCL) prior; the supernode's embedding is mapped through a sigmoid to a five-dimensional confidence vector $\alpha$, so the final estimate is $\hat{x}_{\text{final}} = \alpha \odot \hat{x}_{\text{GNN}} + (1-\alpha) \odot \hat{x}_{\text{WCL}}$. The training loss $L_{\text{CAGE}} = \tfrac{1}{2}(L_{\text{GNN}} + L_{\text{Adapt}}) + \lambda \sum (1-\alpha)^2$ forces the GNN to learn the position independently while still allowing the confidence mechanism to lean on WCL when sampling is dense and symmetric.

What would settle it

Collect a measured dataset with real transceivers at known positions and a known jammer, using the same noise-floor sampling and topology conditions as the static simulations, and compare CAGE against GAT and WCL; if CAGE's RMSE advantage, especially for jammers outside the sampled area, does not reproduce, the central transfer claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a jammer's position can be estimated directly from a graph of noise-floor measurements, and that this graph-based formulation with confidence-guided fusion is more accurate and more robust than classical methods and plain graph neural networks. Across the paper's simulated evaluations, CAGE reaches an overall RMSE of 66.6 m in static scenarios versus 74.4 m for GAT and 143.1 m for WCL, and 15.7 m versus 21.3 m for GAT in dynamic scenarios. The improvement is largest when the jammer is outside the sampled region or signal information is sparse and obfuscated, conditions where WCL and path-loss-based methods degrade sharply. The paper also argues that the explicit GNN supervision term is necessary: without it, the GNN regressor degenerates into a residual corrector for the WCL prior.

Load-bearing premise

The paper's results depend on the assumption that the simulated radio environment used to generate all its data, with signal strength dropping by log-distance path loss plus random shadowing, matches real jamming conditions closely enough for a model trained on those graphs to keep working in the field.

Editorial extensions

If this is right

A trained CAGE model can estimate jammer position on measurement layouts it never saw, because the regression is inductive over graphs rather than tied to fixed node positions.
In the paper's static simulations, CAGE lowers overall RMSE to 66.6 m from 143.1 m for WCL and from 74.4 m for GAT, with the largest margin when the jammer lies outside the sampled region.
In dynamic encirclement trajectories, CAGE's mean RMSE is 15.7 m versus 21.3 m for GAT, and it stays accurate at long ranges where WCL collapses to 372.4 m.
The learned confidence weight makes the system switch behavior: GNN extrapolation dominates at long distance, while the WCL prior takes over near the jammer where sampling is dense and radially symmetric.
The ablations identify reusable design rules for this task: directed KNN edges with small k, max pooling, DropNode at 0.2, and local noise statistics as node features all improve accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper does not test is to apply the same graph representation to other radio sources, such as a malfunctioning transmitter or an unlicensed interferer, since the inputs are only noise-floor values with positions, not jammer-specific features.
The five-dimensional confidence vector could double as a trust signal: a low alpha indicates the estimate leans on the prior and may warrant extra sampling, which the paper does not exploit.
Because the evaluation is entirely simulated under one propagation model family, the decisive test would be a train/test split across different propagation conditions, such as different shadowing or path-loss exponents, which the paper's parameter sweeps suggest but do not run as an explicit transfer experiment.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read it. This is the first GNN application to jamming source localization I know of from the cited literature, and the CAGE architecture is a genuinely new combination: a supernode encoding a WCL prior, plus per-coordinate learned confidence that fuses the GNN prediction with that prior. The paper does two things well. The ablation study is thorough—graph construction, pooling, augmentations, downsampling—and Table 3 is the key evidence: with the L_GNN loss, the GNN regressor alone gets 16.2 m RMSE versus 153.0 without it, so the WCL prior is not secretly carrying the model. The dynamic trajectory results are also clean and consistent across distance intervals.

The soft spots are mostly where the reader put them. All empirical support comes from one synthetic LDPL generator with Gaussian shadowing; there is no out-of-distribution test, no real-world RSSI data, and the data-generation filter (at least one node above -80 dBm, at least three jammed nodes) removes the sparsest and potentially hardest cases. The authors explicitly scope this in Section 6 as future work, but the abstract says 'significantly outperforms' without the caveat, which is broader than the evidence. Also, the 114.9% improvement statement is misleading—it looks like they mean WCL's error is 114.9% higher than CAGE's, not that accuracy improved by that amount. There are no statistical tests on the RMSE differences; with three seeds the standard deviations are small, but formal tests would be better. Finally, CAGE-specific hyperparameters are missing from Table 10—we get them for MLP, GCN, PNA, and GAT, but not for the proposed method, which is odd for reproducibility.

None of these are load-bearing flaws for the claim as a simulation study. The core result—CAGE beats GAT and WCL across static and dynamic scenarios in this environment—holds up on the evidence presented. The issues are about scope generalization and presentation, not internal contradiction.

This paper deserves a serious referee. I'd send it to the wireless-security or GNN-for-localization community. It's a solid new application with a reusable confidence-fusion pattern, and the code/data links are real. I'd ask the authors to add statistical tests, disclose the missing hyperparameters, and include at least one cross-domain sanity check—e.g., a different shadowing distribution or a small real-world RSSI dataset—before publication. Even if that last one doesn't fully converge, the simulation study stands on its own.

Referee Report

3 major / 5 minor

Summary. The paper presents CAGE, a graph neural network framework for localizing a jamming source from noise-floor measurements collected by static or mobile sensors. The authors model each measurement as a node in a KNN graph, use attention-based message passing to aggregate spatial and signal information, and introduce a supernode that encodes a weighted-centroid (WCL) prior. The final position is an element-wise confidence-weighted blend of the GNN prediction and the WCL prior, with the confidence vector produced from the supernode representation. Training uses a composite loss that includes both a direct GNN regression term and the blended adaptive term. The method is evaluated on synthetic LDPL-generated static and dynamic datasets against classical baselines (WCL, PL, MLE, MLAT, LSQ) and learning-based baselines (MLP, GCN, PNA, GAT). The reported results show that CAGE achieves lower RMSE than all baselines in both settings, and ablations identify the contribution of the supernode, confidence outputs, loss terms, graph construction, pooling, and augmentations.

Significance. If the reported gains hold beyond the synthetic evaluation, this is a useful contribution to wireless security: it is, to my knowledge, the first GNN-based treatment of jamming source localization, and it convincingly demonstrates in simulation that a learned graph representation can outperform classical geometric methods, especially in sparse and extrapolation-dominated scenarios. The design separating the GNN regressor from the WCL prior, enforced through the L_GNN loss, is sound and is supported by the ablation in Table 3, where the GNN alone achieves 16.2 m RMSE when trained with L_GNN versus 153.0 m without it. The paper also provides code and data links and includes extensive ablations on graph construction, pooling, augmentations, and downsampling, which strengthen the empirical analysis. The main limitation is that all experiments use a single synthetic LDPL generator, so the practical significance depends on transferability that is not yet demonstrated.

major comments (3)

[Section 6 and Appendix D] The entire evaluation is confined to one synthetic data distribution: LDPL path loss with Gaussian shadowing, the noise-floor formula in Appendix D, and the specific parameter ranges for gamma, sigma, and node counts. Section 6 lists 'evaluating cross-domain generalization beyond simulated environments' as future work, yet the abstract and Section 5.1 state without qualification that the framework 'significantly outperforms established localization baselines.' Real RF environments include correlated shadowing, multipath, non-Gaussian fading, antenna-gain uncertainty, and calibration offsets, none of which are in the generator. This is the load-bearing assumption behind the practical claim, and it is currently unsupported. I recommend either adding a cross-domain experiment (e.g., a different propagation model, ray tracing, or a real measurement campaign) or revising the claims to be explicitly about the simulated LDPL setting.
[Section 5, classical baselines] The implementation and parameter-estimation details for the classical baselines are not given. In particular, path-loss-based methods (PL, MLE, MLAT, LSQ) require estimates of the path-loss exponent gamma and jammer transmit power P_jam^t; the paper attributes their poor performance to estimation error (Section 5.1) but never states how these parameters were estimated or initialized for the comparison. Without this information, the fairness of the baseline comparison cannot be assessed, and this directly affects the central claim of outperforming established baselines. Please specify the exact implementations, any parameter-estimation procedures, and whether the baselines were tuned.
[Tables 1 and 2 and statistical significance] The paper reports results averaged over three seeds but does not provide standard deviations in Tables 1 and 2, and no statistical significance tests are performed. The static-table differences (66.6 m for CAGE vs 74.4 m for GAT) and dynamic-table differences (15.7 vs 21.3) may be meaningful, but with only three trials and no variability measure for the static case, the claim of 'significantly outperforming' is not statistically supported. Please report per-seed results or error bars for the main tables and, if possible, a paired significance test across the three trials.

minor comments (5)

[Section 5.2] The sentence 'Table 9 presents the jammer localization performance for various methods in the dynamic experiment for MAE. RMSE results are provided in Appendix B Table 2' is inconsistent: Table 9 is the MAE table in Appendix B, and Table 2 is the RMSE table in the main text. Please correct the cross-reference.
[Section 5.3, downsampling paragraph] The text says 'spatial boning with noise filtering' where it should say 'spatial binning with noise filtering'.
[Equation (1)] The rendered formula for the edge weight w_ij appears garbled; please check the LaTeX and ensure the intended expression is printed correctly.
[Section 2.1] The sentence 'leveraging the Log Distance Path Loss (LDPL) [6] model' cites reference [6] (Metagraphloc), but the LDPL model is defined later with reference [18] (Rappaport). Please use the correct citation.
[Table 3] The table caption mentions symbols (a square, circle, and triangle) for pooled graph representations, but these symbols are missing from the rendered text, making the table hard to interpret.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CAGE's reported gains come from a genuinely trained GNN evaluated on held-out synthetic instances; the main limitation is simulation-to-reality transfer, not a self-referential derivation.

full rationale

The derivation chain is self-contained. The task is defined as a supervised graph regression problem: nodes and edges are built from the measured noise floor and positions (Eqs. 1–6), the GNN produces an independent position estimate (Eq. 9), and the final output is a learned convex combination of that estimate and the WCL prior (Eq. 12). The paper explicitly guards against the GNN becoming a residual corrector for WCL by adding the LGNN loss (Eq. 15), and the ablation confirms this design works: training with LCAGE yields a GNN-only RMSE of 16.2 m versus 153.0 m without it, while the final adaptive estimate reaches 15.7 m (Table 3). CAGE therefore does not reduce to WCL by construction; WCL is an input component, not the output itself. All reported errors are genuine out-of-sample numbers: models are trained on one set of generated instances and tested on held-out instances from the same LDPL generator described in Appendix D. No fitted parameter is renamed as a prediction, no load-bearing self-citation appears in the derivation, and no uniqueness theorem is imported from the authors. The acknowledged limitation—Section 6 lists 'evaluating cross-domain generalization beyond simulated environments' as future work—is an external-validity concern about the synthetic LDPL data, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the simulated data distribution and the chosen graph architecture. No physical entities are invented; the supernode is a computational construct. The numeric free parameters are hyperparameters the authors selected through ablations, plus the undisclosed CAGE-specific hyperparameters, which are a reproducibility gap.

free parameters (6)

k (KNN neighborhood size) = 3
Chosen via ablation (Table 4a); smaller k reduces oversmoothing and gives best RMSE. Affects graph connectivity that the central results depend on.
Global pooling operation = max
Chosen via ablation (Table 4b); max pooling outperforms sum, mean, and attention pooling.
DropNode rate p = 0.2
Chosen via augmentation ablation (Table 7); DropNode at 0.2 gives best RMSE (52.9) and is used during training only.
CAGE dropout = 0
Hand-tuned; the paper states dropout is set to 0 for CAGE because it gave the best results (Appendix C).
lambda in LCAGE = 0
Set to 0 for simplicity in all experiments; the paper notes it can be tuned, but no sensitivity analysis is given (Section 4.3, Section 5.3).
CAGE architecture hyperparameters = not reported
Number of layers, attention heads, hidden channels, learning rate, batch size, and weight decay for CAGE are not listed in Table 10, so the exact model used for the central results is unspecified.

assumptions (5)

domain assumption The LDPL model with path-loss exponents 2.7-5.0 and shadowing 2-6 dB accurately models NLOS jamming signal propagation
Invoked throughout the synthetic data generation (Appendix D, Eq. 16); if real propagation differs, the simulation results do not transfer.
domain assumption Noise floor at each node is the sum of ambient noise (-100 dBm) and jammer RSSI, and each node knows its own position and measures the noise floor
This defines the input features (Section 3, Appendix D). In a real jamming attack, GPS may be unavailable and the jammer's power/gain may be unknown.
domain assumption KNN graph with exponentially decaying edge weights (Eq. 1) encodes the spatial attenuation principle of jamming signals
This is a modeling choice that shapes message passing; it is not derived from physical principles.
standard math The 5-tuple angular representation (r, sin theta, cos theta, sin phi, cos phi) uniquely represents a position and permits convex blending
Used in Equations (9) and (12); the mapping from Cartesian to this representation is invertible under standard conventions.
domain assumption WCL prior is reliable when node density is high and placement radially symmetric, based on [25]
The confidence mechanism leans on this prior in dense/symmetric conditions (Section 4.2); if the prior were poor everywhere, the blend would still rely on the GNN, but the claimed benefit would shrink.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Graph Neural Networks for Jamming Source Localization." pith.science (2026). https://pith.science/paper/X6DS5RG3

@misc{pith2026250603196,
  author       = {Pith},
  title        = {Pith review of: Graph Neural Networks for Jamming Source Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X6DS5RG3}},
  note         = {Machine review of arXiv:2506.03196}
}

read the original abstract

Graph-based learning provides a powerful framework for modeling complex relational structures; however, its application within the domain of wireless security remains significantly underexplored. In this work, we introduce the first application of graph-based learning for jamming source localization, addressing the imminent threat of jamming attacks in wireless networks. Unlike geometric optimization techniques that struggle under environmental uncertainties and dense interference, we reformulate the localization as an inductive graph regression task. Our approach integrates structured node representations that encode local and global signal aggregation, ensuring spatial coherence and adaptive signal fusion. To enhance robustness, we incorporate an attention-based \ac{GNN} that adaptively refines neighborhood influence and introduces a confidence-guided estimation mechanism that dynamically balances learned predictions with domain-informed priors. We evaluate our approach under complex \ac{RF} environments with various sampling densities, network topologies, jammer characteristics, and signal propagation conditions, conducting comprehensive ablation studies on graph construction, feature selection, and pooling strategies. Results demonstrate that our novel graph-based learning framework significantly outperforms established localization baselines, particularly in challenging scenarios with sparse and obfuscated signal information. Our code is available at https://github.com/tiiuae/gnn-jamming-source-localization.

Figures

Figures reproduced from arXiv: 2506.03196 by the authors.

**Figure 2.** Localization performance as a function of (a) minimum distance to the jammer and (b) maximum noise floor. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Localization accuracy along a single trajectory. The red cross-mark is the jammer position, orange is WCL, and [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Effect of training loss on confidence weighting in CAGE. (a) Training with [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: RMSE analysis of the CAGE, GAT and WCL models across all topologies under the LDPL environment with [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 26 canonical work pages

[1]

In: 2019 IEEE National Aerospace and Electronics Conference (NAECON)

Aldosari, W., Zohdy, M., Olawoyin, R.: Jammer localization through smart estimation of jammer’s transmission power. In: 2019 IEEE National Aerospace and Electronics Conference (NAECON). pp. 430–436 (2019)

work page 2019
[2]

In: International conference on machine learning

Chen, D., O’Bray, L., Borgwardt, K.: Structure-aware transformer for graph representation learning. In: International conference on machine learning. pp. 3469–3489. PMLR (2022) 12 D. Herzalla et al

work page 2022
[3]

In: Advances in Neural Information Processing Systems (2020)

Corso, G., Cavalleri, L., Beaini, D., Liò, P., Veličković, P.: Principal neighbourhood aggregation for graph nets. In: Advances in Neural Information Processing Systems (2020)

work page 2020
[4]

ACM SIGKDD Explorations Newsletter24(2), 61–77 (2022)

Ding, K., Xu, Z., Tong, H., Liu, H.: Data augmentation for deep graph learning: A survey. ACM SIGKDD Explorations Newsletter24(2), 61–77 (2022)

work page 2022
[5]

Expert Systems with Applications174, 114711 (2021)

Do, T.H., Nguyen, D.M., Bekoulis, G., Munteanu, A., Deligiannis, N.: Graph convolutional neural networks with node transition probability-based message passing and dropnode regularization. Expert Systems with Applications174, 114711 (2021)

work page 2021
[6]

arXiv preprint arXiv:2411.17781 (2024)

Etiabi, Y., Eldeeb, E., Shehab, M., Njima, W., Alves, H., Alouini, M.S., Amhoud, E.M.: Metagraphloc: A graph-based meta-learning scheme for indoor localization via sensor fusion. arXiv preprint arXiv:2411.17781 (2024)

arXiv 2024
[7]

International Journal of Antennas and Propagation2019(1), 1523469 (2019)

Häfner, S., Käske, M., Thomä, R.: On calibration and direction finding with uniform circular arrays. International Journal of Antennas and Propagation2019(1), 1523469 (2019)

work page 2019
[8]

In: International Conference on Learning Representations (2017)

Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (2017)

2017

Show all 31 references

[9]

In: Machine Learning for Indoor Localization and Navigation, pp

Lezama, F., Larroca, F., Capdehourat, G.: On the application of graph neural networks for indoor positioning systems. In: Machine Learning for Indoor Localization and Navigation, pp. 239–256. Springer (2023)

2023
[10]

In: Proceedings of the AAAI conference on artificial intelligence

Li, Q., Han, Z., Wu, X.M.: Deeper insights into graph convolutional networks for semi-supervised learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)

2018
[11]

In: 2009 IEEE International Conference on Pervasive Computing and Communications

Liu, H., X, W., Chen, Y., Liu, Z.: Localizing jammers in wireless networks. In: 2009 IEEE International Conference on Pervasive Computing and Communications. pp. 1–6 (2009)

2009
[12]

Nature624(7990), 80–85 (2023)

Merchant,A.,Batzner,S.,Schoenholz,S.S.,Aykol,M.,Cheon,G.,Cubuk,E.D.:Scalingdeeplearningformaterialsdiscovery. Nature624(7990), 80–85 (2023)

2023
[13]

In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Nardin, A., Imbiriba, T., Closas, P.: Jamming source localization using augmented physics-based model. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5 (2023)

2023
[15]

In: 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC)

Niu, Z., Li, H., Zhou, X., Huang, J.: Overview of jammer localization in wireless sensor networks. In: 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC). vol. 9, pp. 9–13 (2020)

2020
[16]

In: GLOBECOM 2009 - 2009 IEEE Global Telecommunications Conference

Pelechrinis, K., Koutsopoulos, I., Broustis, I., Krishnamurthy, S.V.: Lightweight jammer localization in wireless networks: System design and implementation. In: GLOBECOM 2009 - 2009 IEEE Global Telecommunications Conference. pp. 1–6 (2009)

2009
[17]

IEEE Communications Surveys & Tutorials24(2), 767–809 (2022)

Pirayesh, H., Zeng, H.: Jamming attacks and anti-jamming strategies in wireless networks: A comprehensive survey. IEEE Communications Surveys & Tutorials24(2), 767–809 (2022)

2022
[18]

Cambridge University Press (1996)

Rappaport, T.S.: Wireless communications: principles and practice. Cambridge University Press (1996)

1996
[19]

In: Proceedings of the 2021 SIAM international conference on data mining (SDM)

Sato, R., Yamada, M., Kashima, H.: Random features strengthen graph neural networks. In: Proceedings of the 2021 SIAM international conference on data mining (SDM). pp. 333–341. SIAM (2021)

2021
[20]

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res.15(1), 1929–1958 (Jan 2014)

1929
[21]

Sensors22(01 2022)

Tedeschi, P., Oligeri, G., Pietro, R.: Localization of a power-modulated jammer. Sensors22(01 2022)

2022
[22]

In: 2021 International Conference on 3D Vision (3DV)

Turkoglu, M.O., Brachmann, E., Schindler, K., Brostow, G.J., Monszpart, A.: Visual camera re-localization using graph neural networks and relative pose supervision. In: 2021 International Conference on 3D Vision (3DV). pp. 145–155 (2021)

2021
[23]

Current Opinion in Structural Biology79, 102538 (2023)

Veličković, P.: Everything is connected: Graph neural networks. Current Opinion in Structural Biology79, 102538 (2023)

2023
[24]

In: International Conference on Learning Representations (2018)

Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., Bengio, Y.: Graph attention networks. In: International Conference on Learning Representations (2018)

2018
[25]

IEEE Transactions on wireless communications10(10), 3403–3413 (2011)

Wang, J., Urriza, P., Han, Y., Cabric, D.: Weighted centroid localization algorithm: Theoretical analysis and distributed implementation. IEEE Transactions on wireless communications10(10), 3403–3413 (2011)

2011
[26]

arXiv preprint arXiv:1909.01315 (2019)

Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y., Xiao, T., He, T., Karypis, G., Li, J., Zhang, Z.: Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315 (2019)

2019 arXiv
[27]

arXiv preprint arXiv:2009.10564 (2020)

Wang, Y., Wang, W., Liang, Y., Cai, Y., Hooi, B.: Graphcrop: Subgraph cropping for graph classification. arXiv preprint arXiv:2009.10564 (2020)

2020 arXiv
[28]

IEEE Communications Surveys & Tutorials19(2), 765–799 (2017)

Wei, X., Wang, Q., Wang, T., Fan, J.: Jammer localization in multi-hop wireless network: A comprehensive survey. IEEE Communications Surveys & Tutorials19(2), 765–799 (2017)

2017
[29]

arXiv preprint arXiv:2311.16856 (2023)

Yan, W., Wang, J., Yin, F., Tian, Y., Zoubir, A.M.: Attentional graph neural networks for robust massive network local- ization. arXiv preprint arXiv:2311.16856 (2023)

2023 arXiv
[30]

In: 2016 International Conference on Indoor Positioning and Indoor Navigation (IPIN)

Yang, J., Lee, H., Moessner, K.: Multilateration localization based on singular value decomposition for 3d indoor positioning. In: 2016 International Conference on Indoor Positioning and Indoor Navigation (IPIN). pp. 1–8 (2016)

2016
[31]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Yang, Y., Ren, Z., Li, H., Zhou, C., Wang, X., Hua, G.: Learning dynamics via graph neural networks for human pose estimation and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8074–8084 (2021)

2021
[32]

LR” denotes learning rate, “WD

Zhang, M., Li, P.: Nested graph neural networks. Advances in Neural Information Processing Systems34, 15734–15747 (2021) Graph Neural Networks for Jamming Source Localization 13 A Ablation Studies We conduct a series of ablation studies to evaluate the impact of different desi...

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

In: 2019 IEEE National Aerospace and Electronics Conference (NAECON)

Aldosari, W., Zohdy, M., Olawoyin, R.: Jammer localization through smart estimation of jammer’s transmission power. In: 2019 IEEE National Aerospace and Electronics Conference (NAECON). pp. 430–436 (2019)

work page 2019

[2] [2]

In: International conference on machine learning

Chen, D., O’Bray, L., Borgwardt, K.: Structure-aware transformer for graph representation learning. In: International conference on machine learning. pp. 3469–3489. PMLR (2022) 12 D. Herzalla et al

work page 2022

[3] [3]

In: Advances in Neural Information Processing Systems (2020)

Corso, G., Cavalleri, L., Beaini, D., Liò, P., Veličković, P.: Principal neighbourhood aggregation for graph nets. In: Advances in Neural Information Processing Systems (2020)

work page 2020

[4] [4]

ACM SIGKDD Explorations Newsletter24(2), 61–77 (2022)

Ding, K., Xu, Z., Tong, H., Liu, H.: Data augmentation for deep graph learning: A survey. ACM SIGKDD Explorations Newsletter24(2), 61–77 (2022)

work page 2022

[5] [5]

Expert Systems with Applications174, 114711 (2021)

Do, T.H., Nguyen, D.M., Bekoulis, G., Munteanu, A., Deligiannis, N.: Graph convolutional neural networks with node transition probability-based message passing and dropnode regularization. Expert Systems with Applications174, 114711 (2021)

work page 2021

[6] [6]

arXiv preprint arXiv:2411.17781 (2024)

Etiabi, Y., Eldeeb, E., Shehab, M., Njima, W., Alves, H., Alouini, M.S., Amhoud, E.M.: Metagraphloc: A graph-based meta-learning scheme for indoor localization via sensor fusion. arXiv preprint arXiv:2411.17781 (2024)

arXiv 2024

[7] [7]

International Journal of Antennas and Propagation2019(1), 1523469 (2019)

Häfner, S., Käske, M., Thomä, R.: On calibration and direction finding with uniform circular arrays. International Journal of Antennas and Propagation2019(1), 1523469 (2019)

work page 2019

[8] [8]

In: International Conference on Learning Representations (2017)

Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (2017)

2017