REVIEW 3 major objections 4 minor 29 references
LHGNN: Local-Higher Order Graph Neural Networks For Audio Classification and Tagging
T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read A graph network with local and cluster edges outperforms Transformer audio systems on three benchmarks.
desk verdict A sensible fused local/higher-order graph kernel for audio tagging, but the headline 'beats AST' claim is undercut by comparing LHGNN trained with the PSLA recipe against AST numbers from its own pipeline. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the Local-Higher Order Graph Convolution inside each LHG block. Given a flattened spectrogram feature map $X$, each node $x_i$ computes its k-NN set $S_i$ and, via Fuzzy C-Means clustering, membership scores to $P$ centroids; after updating the centroids, the $K$ centroids with highest membership form the set $L_i$. The node is updated as $x''_i = \sigma(x_i \oplus \max(S_i - x_i) \oplus \max(L_i - x_i))$, where $\sigma$ is an MLP with GELU non-linearity, and the result is projected back to the original dimension and added residually. The max-relative operations compare each node with its neighbours and with its cluster centroids along every feature dimension, and the concatenation fuses local and higher-order evidence into one vector; the paper shows this fusion is what lifts performance.
What would settle it
Retrain AST and LHGNN on AudioSet with the identical training pipeline and compare mAP with confidence intervals; if AST matches or exceeds LHGNN's 0.466, the paper's outperformance claim fails.
Extended reading notes
Core claim
LHGNN's central discovery is that fusing complementary local and higher-order graph relations lets a comparatively small graph network match or exceed transformer performance on audio tagging and classification. On AudioSet, LHGNN reaches 0.466 mAP with 31M parameters against AST's 0.459 with 87M; without ImageNet pretraining the gap widens in its favour (0.442 vs 0.366). On FSD50K it reaches 0.59 mAP with pretraining and 0.573 without, versus 0.574 for pretrained AST, while on ESC50 it reaches 96.2% accuracy versus 95.6% for AST. The paper attributes the gain to its proposed graph convolution, which concatenates the node with max-relative differences to local neighbours and to the top Fuzzy C-Means centroids, and its ablations show that each branch alone is weaker than the combination.
Load-bearing premise
The central claim depends on comparing LHGNN's numbers with published Transformer baselines that were trained under different pipelines, with no error bars or significance tests, so if those baselines were re-run under LHGNN's exact training protocol the reported margins could vanish or reverse.
Editorial extensions
If this is right
- If the reported AudioSet numbers hold, LHGNN beats AST by 0.466 to 0.459 mAP while using 31M parameters instead of 87M, making it a parameter-efficient alternative for weakly labelled sound event tagging.
- Without ImageNet pretraining, LHGNN keeps 0.442 mAP on AudioSet and 0.573 on FSD50K, versus 0.366 and 0.396 for AST, so the architecture's advantage is largest precisely when pretrained backbones are unavailable.
- The FSD50K ablation shows the combined kernel (0.573 mAP) outperforms the local-only kernel (0.531) and the cluster-only kernel (0.501), indicating that fusing both relation types is necessary for the reported gain.
- With ImageNet pretraining LHGNN reaches top scores on FSD50K (0.59 mAP) and ESC50 (96.2% accuracy), so the model remains competitive in data-rich settings rather than only in data-scarce ones.
Reading between the lines
- A natural untested extension is to apply the local-plus-cluster kernel to other spectrogram tasks such as music tagging or speech emotion recognition, where ImageNet pretraining is often unavailable and the reported pretraining-free gains would matter most.
- The ablation's density-based clustering score (0.574) edges out Fuzzy C-Means (0.573), so a computationally efficient variant of density-based clustering might give a further gain; the paper does not test that.
- Because LHGNN needs 30 epochs on AudioSet versus 5 for pretrained AST, parameter efficiency does not automatically mean compute efficiency; a fair comparison would also include total training FLOPs and wall-clock time.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper introduces LHGNN, a pyramid graph neural network for audio classification and tagging. Each LHG block updates tokens using two complementary relation types: k-NN local neighborhoods and Fuzzy C-Means cluster centers, fusing them with a max-relative graph convolution followed by a ConvFFN. The model is evaluated on AudioSet, FSD50K, and ESC50. The authors report that LHGNN surpasses AST accuracies while using fewer parameters, particularly without ImageNet pretraining, and they provide ablations on FSD50K supporting the combined kernel and fuzzy clustering.
Significance. If the comparison were controlled, the result would be a useful data point: a 31M-parameter GNN-style model competitive with an 87M-parameter transformer, with better from-scratch scores on AudioSet and FSD50K. The ablations in Tables IV and V are clearly presented and support the design choices of the combined kernel and Fuzzy C-Means clustering. However, the current evidence is not sufficient to support the headline claim because baseline scores are quoted from prior papers rather than re-trained under the same pipeline, and the main margins are small with no error bars. The paper does not provide code or pretrained models, limiting reproducibility, though this is not the central issue.
major comments (3)
- [III-A, Tables I and II] The central comparison to AST is confounded by training pipeline. LHGNN is trained with the PSLA pipeline (mixup, SpecAugment, checkpoint averaging) as stated in Section III-A.1, while AST's numbers are quoted from its original paper and use its own pipeline. On AudioSet the gap is 0.007 mAP (0.466 vs 0.459), and on FSD50K it is 0.016 (0.59 vs 0.574); these margins are small relative to typical seed variance. To support the claim that LHGNN outperforms transformer-based models, the authors should retrain AST (and ideally PSLA) under the identical training recipe, and report means and standard deviations over multiple seeds.
- [II-C] The hyperparameters k and K are selected for best results and changed when ImageNet pretraining is used (k=25/K=10 without, k=9/K=5 with), but the manuscript does not describe a validation protocol or sensitivity analysis. Since the final comparisons hinge on these choices, and since k, K, and P are in effect free parameters, the paper should either fix them a priori or report robustness across ranges.
- [III-C.2] The claim of a 'distinct advantage in scenarios lacking ImageNet pretraining' is not supported on ESC50: LHGNN is only evaluated with ImageNet pretraining (96.2), and the no-pretraining baseline ERANN reaches 96.1, which the text itself acknowledges as 'equally well'. Please either add an LHGNN-from-scratch ESC50 result or soften the claim to be specific to AudioSet and FSD50K.
minor comments (4)
- [II-B1, Eq. (3)] Calling the operation in Eq. (3) a 'graph convolution' is misleading because no learned weights appear inside the max-relative aggregation; the only learnable projection is h(·) after concatenation. Please clarify or rename the operation.
- [III-B] The statement that LHGNN without pretraining is 'comparable to AST with ImageNet pretraining' is based on a 0.001 mAP difference (0.573 vs 0.574), which is within noise; please add variance estimates or rephrase the claim.
- [IV, Table IV caption] The caption says 'different kernels', but the three rows are different input features to the same kernel; 'different input components' would be more accurate.
- [II-B1] The k-NN step does not specify the distance metric; please state whether Euclidean distance is used.
Circularity Check
No significant circularity: the paper's claims are empirical comparisons against public test sets, and its graph update equations are feed-forward architectural definitions rather than derivations from target metrics.
full rationale
The paper's derivation chain is an architecture proposal plus an empirical evaluation, not a claim that some target quantity is derived from the model's own inputs. The LHG update in Eqs. (3)-(4) defines the updated node yi in terms of k-NN sets Si and Fuzzy C-Means centroid sets Li computed from the same input X, but this is a standard feed-forward construction: the clustering is part of the model's computation, and the reported mAP and accuracy numbers come from public held-out test sets (AudioSet evaluation set, FSD50K evaluation set, ESC-50 folds). Thus the target metric is not an input to the model and no fitted parameter is renamed as a prediction. The hyperparameters k, K, and P are tuned, and the paper openly states that the training pipeline follows PSLA [20] and that scores are computed with checkpoint averaging; even if the quoted AST baseline numbers come from a different training pipeline, this is an experimental comparability limitation, not circular reasoning. The only self-citation, ATGNN [11], appears as related work and is not load-bearing for any claim. No equation in the paper is defined in terms of its own predicted output, and no theoretical result is imported from a self-citation chain to force the model choice. The evaluation is also externally anchored to public benchmarks, so the most common circularity failure modes are absent. Score 0 is appropriate.
Assumptions & free parameters
free parameters (5)
- k (kNN neighbors) =
25 without pretraining; 9 with ImageNet pretraining
- K (top centroids per node) =
10 without pretraining; 5 with pretraining
- P (number of FCM centroids) =
50
- Fuzziness parameter m =
2
- FCM iterations v =
1
assumptions (3)
- ad hoc to paper max(Si - xi) and max(Li - xi) capture complementary local and higher-order relations when concatenated
- domain assumption A single FCM iteration (v=1) produces useful centroids within every block
- domain assumption Baseline results from prior papers are directly comparable to LHGNN numbers despite different training pipelines
Cite this review
Pith. "Pith review of LHGNN: Local-Higher Order Graph Neural Networks For Audio Classification and Tagging." pith.science (2026). https://pith.science/paper/PL2LHUNO
@misc{pith2026250103464,
author = {Pith},
title = {Pith review of: LHGNN: Local-Higher Order Graph Neural Networks For Audio Classification and Tagging},
year = {2026},
howpublished = {\url{https://pith.science/paper/PL2LHUNO}},
note = {Machine review of arXiv:2501.03464}
}
read the original abstract
Transformers have set new benchmarks in audio processing tasks, leveraging self-attention mechanisms to capture complex patterns and dependencies within audio data. However, their focus on pairwise interactions limits their ability to process the higher-order relations essential for identifying distinct audio objects. To address this limitation, this work introduces the Local- Higher Order Graph Neural Network (LHGNN), a graph based model that enhances feature understanding by integrating local neighbourhood information with higher-order data from Fuzzy C-Means clusters, thereby capturing a broader spectrum of audio relationships. Evaluation of the model on three publicly available audio datasets shows that it outperforms Transformer-based models across all benchmarks while operating with substantially fewer parameters. Moreover, LHGNN demonstrates a distinct advantage in scenarios lacking ImageNet pretraining, establishing its effectiveness and efficiency in environments where extensive pretraining data is unavailable.
Figures
Reference graph
Works this paper leans on
-
[1]
Sound event detection: A tutorial,
Annamaria Mesaros, Toni Heittola, Tuomas Virtanen, and Mark D Plumbley, “Sound event detection: A tutorial,” IEEE Signal Processing Magazine, 2021
work page 2021
-
[2]
Deep learning in music recommendation systems,
Markus Schedl, “Deep learning in music recommendation systems,” Frontiers in Applied Mathematics and Statistics , 2019
work page 2019
-
[3]
Deep spoken keyword spotting: An overview,
Iv ´an L ´opez-Espejo, Zheng-Hua Tan, John HL Hansen, and Jesper Jensen, “Deep spoken keyword spotting: An overview,” IEEE Access , 2021
work page 2021
-
[4]
Rethinking cnn models for audio classification,
Kamalesh Palanisamy, Dipika Singhania, and Angela Yao, “Rethinking cnn models for audio classification,” arXiv preprint arXiv:2007.11154 , 2020
arXiv 2007
-
[5]
Ast: Audio spectrogram transformer,
Yuan Gong, Yu-An Chung, and James Glass, “Ast: Audio spectrogram transformer,” arXiv preprint arXiv:2104.01778 , 2021
arXiv 2021
-
[6]
James Liang, Yiming Cui, Qifan Wang, Tong Geng, Wenguan Wang, and Dongfang Liu, “Clusterfomer,” Advances in Neural Information Processing Systems, vol. 36, 2024
work page 2024
-
[7]
Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao, “Hypergraph neural networks,” in Proceedings of the AAAI conference on artificial intelligence , 2019
work page 2019
-
[8]
Vision hgnn: An image is more than a graph of nodes,
Yan Han, Peihao Wang, Souvik Kundu, Ying Ding, and Zhangyang Wang, “Vision hgnn: An image is more than a graph of nodes,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19878–19888
2023
Show all 29 references
-
[9]
Everything is connected: Graph neural networks,
Petar Veli ˇckovi´c, “Everything is connected: Graph neural networks,” Current Opinion in Structural Biology , 2023
2023
-
[10]
Vision gnn: An image is worth graph of nodes,
Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu, “Vision gnn: An image is worth graph of nodes,” Advances in neural information processing systems , 2022
2022
-
[11]
Atgnn: Audio tagging graph neural network,
Shubhr Singh, Christian J Steinmetz, Emmanouil Benetos, Huy Phan, and Dan Stowell, “Atgnn: Audio tagging graph neural network,” IEEE Signal Processing Letters , 2024
2024
-
[12]
Cmt: Convolutional neural networks meet vision transformers,
Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu, “Cmt: Convolutional neural networks meet vision transformers,” in CVPR, 2022
2022
-
[13]
Gaussian error linear units (gelus),
Dan Hendrycks and Kevin Gimpel, “Gaussian error linear units (gelus),” arXiv preprint arXiv:1606.08415 , 2016
2016 arXiv
-
[14]
Orthogonal transformer,
Huaibo Huang, Xiaoqiang Zhou, and Ran He, “Orthogonal transformer,” Advances in Neural Information Processing Systems , 2022
2022
-
[15]
How do vision transformers work?,
Namuk Park and Songkuk Kim, “How do vision transformers work?,” ICLR, 2022
2022
-
[16]
Srformer: Permuted self-attention for single image super-resolution,
Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou, “Srformer: Permuted self-attention for single image super-resolution,” in CVPR, 2023
2023
-
[17]
Audioset,
Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter, “Audioset,” in ICASSP, 2017
2017
-
[18]
Eduardo Fonseca, Xavier Favory, Jordi Pons, Frederic Font, and Xavier Serra, “Fsd50k,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021
2021
-
[19]
Esc-50: Dataset for environmental sound classification,
Karol J Piczak, “Esc-50: Dataset for environmental sound classification,” in Proceedings of the 23rd ACM international conference on Multimedia, 2015, pp. 1015–1018
2015
-
[20]
PSLA: Improving audio tagging with pretraining, sampling, labeling, and aggregation,
Yuan Gong, Yu-An Chung, and James Glass, “PSLA: Improving audio tagging with pretraining, sampling, labeling, and aggregation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2021
2021
-
[21]
mixup: Beyond empirical risk minimization,
Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez- Paz, “mixup: Beyond empirical risk minimization,” arXiv preprint arXiv:1710.09412, 2017
2017 arXiv
-
[22]
Specaugment: A simple data augmentation method for automatic speech recognition,
Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” arXiv preprint arXiv:1904.08779, 2019
1904 arXiv
-
[23]
A deep residual network for large-scale acoustic scene analysis.,
Logan Ford, Hao Tang, Franc ¸ois Grondin, and James R Glass, “A deep residual network for large-scale acoustic scene analysis.,” inInterSpeech, 2019
2019
-
[24]
Panns: Large-scale pretrained audio neural networks for audio pattern recognition,
Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley, “Panns: Large-scale pretrained audio neural networks for audio pattern recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2020
2020
-
[25]
Wav2clip: Learning robust audio representations from clip,
Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello, “Wav2clip: Learning robust audio representations from clip,” in ICASSP. IEEE, 2022
2022
-
[26]
Audio transformers,
Prateek Verma and Jonathan Berger, “Audio transformers,” arXiv preprint arXiv:2105.00335, 2021
2021 arXiv
-
[27]
Learning transferable visual models from natural language supervision,
Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning....
2021
-
[28]
Sergey Verbitskiy, Vladimir Berikov, and Viacheslav Vyshegorodtsev, “Eranns,” Pattern Recognition Letters, 2022
2022
-
[29]
ESC: Dataset for Environmental Sound Classification,
Karol J. Piczak, “ESC: Dataset for Environmental Sound Classification,” in Proceedings of the 23rd Annual ACM Conference on Multimedia
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.