pith. sign in

arxiv: 2605.25061 · v1 · pith:2G7JKKBOnew · submitted 2026-05-24 · 💻 cs.LG · cs.AI

GL-LFGNN:A Global-Local Dual-branch Causal Graph Neural Network Based on Liang-Kleeman Information Flow for EEG Emotion Recognition

Ziyi Wang , Dongyang Kuang This is my paper

Pith reviewed 2026-06-30 12:18 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords EEG emotion recognitioncausal graph neural networksLiang-Kleeman information flowglobal-local dual-brancharousal valence classificationMEEG datasetdirected graphslow-parameter models
0
0 comments X

The pith

A dual-branch graph neural network using directed causal information flow from Liang-Kleeman theory reaches 86 percent accuracy on EEG emotion tasks with only 37K parameters.

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

The paper seeks to replace symmetric correlation-based graphs in EEG emotion recognition models with directed graphs that capture actual causal influences between brain channels. It builds GL-LFGNN around Liang-Kleeman information flow to quantify these influences from a dynamical systems view and adds a dual-branch structure that processes whole-brain and region-specific connectivity separately. The resulting network delivers high accuracy on arousal and valence classification while using roughly one-tenth the parameters of prior leading methods, which suggests that aligning the model with the directed nature of neural signaling can improve both performance and efficiency at once.

Core claim

GL-LFGNN constructs directed adjacency matrices from Liang-Kleeman information flow to model asymmetric causal influences among EEG channels, then processes them through a global-local dual-branch architecture aligned with functional neuroanatomy. This replaces the symmetric adjacency matrices used in earlier GNNs and produces a model that reaches 86.17 percent accuracy on arousal and 86.71 percent on valence in the MEEG dataset while requiring only 37K parameters, about 10 percent of current state-of-the-art size.

What carries the argument

Liang-Kleeman information flow, which quantifies directed causal strength between time series from a dynamical systems perspective to generate neurophysiologically interpretable directed graphs for the neural network.

If this is right

  • Accuracy reaches 86.17 percent for arousal and 86.71 percent for valence on the MEEG dataset.
  • Parameter count stays at 37K, roughly 10 percent of existing state-of-the-art models.
  • Directed causal graphs improve neurophysiological interpretability over correlation matrices.
  • Global-local dual-branch processing combines whole-brain connectivity with region-specific features.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same causal-graph construction could be tested on other EEG classification tasks such as motor imagery or sleep staging.
  • If the efficiency gains hold, the approach may encourage smaller, more deployable models in wearable brain-signal devices.
  • Further validation on multi-site or clinical EEG recordings would be required to establish whether the causal advantage persists outside the MEEG dataset.

Load-bearing premise

Liang-Kleeman information flow theory supplies directed graphs that better represent neural information flow than symmetric matrices built from spatial proximity or correlations.

What would settle it

A symmetric correlation-based GNN matching or exceeding both the 86 percent accuracy figures and the 37K parameter count on the MEEG dataset would undermine the claimed advantage of the causal construction.

Figures

Figures reproduced from arXiv: 2605.25061 by Dongyang Kuang, Ziyi Wang.

Figure 1
Figure 1. Figure 1: The overall framework of GL-LFGNN 2.1 Causal graph construction Liang–Kleeman information flow quantifies directed information transfer be￾tween components of a dynamical system based on entropy variation, and pro￾vides an effective tool for characterizing causal interactions in multivariate time series [20]. Within this framework, a variable Xj is considered causal to Xi if the information flow Tj→i is no… view at source ↗
Figure 2
Figure 2. Figure 2: Brain regionalization 2.3 Information aggregation We constructs Hierarchical Causality Graph Convolutional Neural Network (HCD -GCN) to complete information aggregation, which consists of two directed graph diffusion convolutional (DConv [24]) layers and a graph pooling (DIFFPOOL [25] or Attention Pool) layer.Assuming that the adjacency matrix is A ∈ R n×n and the feature matrix X ∈ R n×d , the out-degree … view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of local branch attention weights witch are the average of 10-fold cross-validation. 4 Conclusion We present GL-LFGNN, which bridges a critical gap between correlational graph learning and the causal nature of neural information flow. By grounding adja￾cency matrices in Liang-Kleeman information flow theory, our model provides both rigorous causal semantics and neurophysiological interpretabi… view at source ↗
read the original abstract

EEG-based emotion recognition holds significant promise for objective diagnosis of mood disorders. Graph neural networks (GNNs) have emerged as the dominant paradigm for modeling inter-channel dependencies in EEG, yet existing approaches rely on symmetric adjacency matrices derived from spatial proximity or functional correlations that fundamentally capture statistical associations rather than directed causal influences, which conflicts with the inherently asymmetric, causally-driven nature of neural information flow. To bridge this gap, we propose GL-LFGNN, a Global-Local Dual-branch Causal Graph Neural Network grounded in Liang-Kleeman information flow theory. Unlike Granger causality that merely assesses temporal precedence, our approach rigorously quantifies causal strength from a dynamical systems perspective, yielding neurophysiologically interpretable directed graphs. A dual-branch architecture further integrates whole-brain connectivity with region-specific processing aligned to established functional neuroanatomy. On the MEEG dataset, GL-LFGNN achieves 86.17% (Arousal) and 86.71% (Valence) accuracy with only 37K parameters -- approximately 10% of the current state-of-the-art -- demonstrating that principled causal modeling can simultaneously enhance interpretability, generalization, and computational efficiency. Code will be released.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes GL-LFGNN, a global-local dual-branch causal graph neural network for EEG emotion recognition that constructs directed adjacency matrices using Liang-Kleeman information flow theory rather than symmetric correlation or spatial matrices. It integrates whole-brain and region-specific branches aligned with neuroanatomy and reports 86.17% arousal and 86.71% valence accuracy on the MEEG dataset using only 37K parameters (approximately 10% of current SOTA), claiming simultaneous gains in interpretability, generalization, and efficiency. Code release is promised.

Significance. If the performance claims hold after proper validation and the contribution of the causal graphs can be isolated from the dual-branch architecture, the work would offer a more principled dynamical-systems approach to modeling directed neural information flow in GNNs for EEG, potentially improving both accuracy and neurophysiological interpretability while reducing model size. The explicit promise of code release is a positive factor for reproducibility.

major comments (2)
  1. [Abstract] Abstract: the central performance claims (86.17% arousal, 86.71% valence accuracy, 37K parameters) are stated without any information on data splits, baseline comparisons, statistical testing, hyperparameter selection, or validation of the constructed causal graphs, so the support for the claim that principled causal modeling drives the results cannot be evaluated.
  2. [Abstract] Abstract: no ablation is reported that replaces the Liang-Kleeman-derived directed causal adjacency with correlation-based or spatial-proximity matrices while holding the global-local dual-branch architecture fixed; without this isolation the attribution of gains specifically to the causal modeling (rather than the dual-branch design) remains untested and load-bearing for the central claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestions. We address each major comment below and indicate the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claims (86.17% arousal, 86.71% valence accuracy, 37K parameters) are stated without any information on data splits, baseline comparisons, statistical testing, hyperparameter selection, or validation of the constructed causal graphs, so the support for the claim that principled causal modeling drives the results cannot be evaluated.

    Authors: While the abstract is designed to be concise, the full manuscript provides comprehensive details on these aspects in the Methods and Experiments sections. Specifically, we use subject-independent 5-fold cross-validation, compare against several state-of-the-art baselines, report statistical significance, describe hyperparameter selection, and validate the causal graphs through neurophysiological interpretability analysis. To improve accessibility, we will revise the abstract to briefly reference the experimental validation protocol. revision: partial

  2. Referee: [Abstract] Abstract: no ablation is reported that replaces the Liang-Kleeman-derived directed causal adjacency with correlation-based or spatial-proximity matrices while holding the global-local dual-branch architecture fixed; without this isolation the attribution of gains specifically to the causal modeling (rather than the dual-branch design) remains untested and load-bearing for the central claim.

    Authors: We agree that an ablation isolating the causal adjacency while fixing the architecture would provide stronger evidence for the contribution of the Liang-Kleeman information flow. Although the paper includes comparisons to alternative adjacency constructions, they are not held exactly fixed to the dual-branch design. We will add this specific ablation experiment in the revised manuscript to directly address this point. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation relies on external theory and empirical reporting

full rationale

The provided abstract and description contain no equations, self-citations, fitted parameters presented as predictions, or load-bearing uniqueness theorems from the authors. The method adopts Liang-Kleeman information flow (an external dynamical-systems result) to construct directed graphs, then adds a dual-branch architecture and reports dataset accuracies. These steps do not reduce by construction to the inputs; the accuracies are empirical measurements, not tautological outputs of the graph-construction rule. No self-definitional, fitted-input, or ansatz-smuggling patterns appear.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review supplies insufficient detail to enumerate concrete free parameters or invented entities beyond the high-level architectural choice; the central modeling assumption is the applicability of Liang-Kleeman theory to neural causality.

axioms (1)
  • domain assumption Liang-Kleeman information flow quantifies directed causal strength from a dynamical systems perspective and produces neurophysiologically interpretable graphs
    Invoked as the foundation for replacing symmetric adjacency matrices
invented entities (1)
  • Global-Local Dual-branch architecture no independent evidence
    purpose: Integrates whole-brain connectivity with region-specific processing aligned to functional neuroanatomy
    New architectural component introduced to combine global and local causal graphs

pith-pipeline@v0.9.1-grok · 5750 in / 1466 out tokens · 34568 ms · 2026-06-30T12:18:01.311178+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages · 2 internal anchors

  1. [1]

    In: 2016 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp

    Li,X.,Song,D.,Zhang,P.,Yu,G.,Hou,Y.,Hu,B.:Emotionrecognitionfrommulti- channel EEG data through Convolutional Recurrent Neural Network. In: 2016 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 352–359. IEEE (2016). https://doi.org/10.1109/BIBM.2016.7822545

    work page doi:10.1109/bibm.2016.7822545 2016

  2. [2]

    Human Brain Mapping38(11), 5391–5420 (2017)

    Schirrmeister, R.T., Springenberg, J.T., Fiederer, L.D.J., Glasstetter, M., Eggensperger, K., Tangermann, M., Hutter, F., Burgard, W., Ball, T.: Deep learn- ing with convolutional neural networks for EEG decoding and visualization. Human Brain Mapping38(11), 5391–5420 (2017). https://doi.org/10.1002/hbm.23730

    work page doi:10.1002/hbm.23730 2017

  3. [3]

    IEEE Trans

    Zhang, T., Zheng, W., Cui, Z., Zong, Y., Li, Y.: Spatial–temporal recurrent neural network for emotion recognition. IEEE Trans. Cybernetics49(3), 839–847 (2019). https://doi.org/10.1109/TCYB.2017.2788081

    work page doi:10.1109/tcyb.2017.2788081 2019

  4. [4]

    Journal of Neural Engineering15(5), 056013 (2018).https: //doi.org/10.1088/1741-2552/aace8c

    Lawhern, V.J., Solon, A.J., Waytowich, N.R., Gordon, S.M., Hung, C.P., Lance, B.J.: EEGNet: a compact convolutional neural network for EEG-based brain– computer interfaces. Journal of Neural Engineering15(5), 056013 (2018).https: //doi.org/10.1088/1741-2552/aace8c

    work page doi:10.1088/1741-2552/aace8c 2018

  5. [5]

    Cognitive Neurodynamics14, 815–828 (2020)

    Shen, F., Dai, G., Lin, G., et al.: EEG-based emotion recognition using 4D con- volutional recurrent neural network. Cognitive Neurodynamics14, 815–828 (2020). https://doi.org/10.1007/s11571-020-09634-1

    work page doi:10.1007/s11571-020-09634-1 2020

  6. [6]

    IEEE Signal Processing Magazine30(3), 83–98 (2013)

    Shuman,D.I.,Narang,S.K.,Frossard,P.,Ortega,A.,Vandergheynst,P.:Theemerg- ing field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE Signal Processing Magazine30(3), 83–98 (2013). https://doi.org/10.1109/MSP.2012.2235192

    work page doi:10.1109/msp.2012.2235192 2013

  7. [7]

    IEEE Trans

    Zhong, P., Wang, D., Miao, C.: EEG-based emotion recognition using regularized graph neural networks. IEEE Trans. Affective Computing13(3), 1290–1301 (2022). https://doi.org/10.1109/TAFFC.2020.2994159

    work page doi:10.1109/taffc.2020.2994159 2022

  8. [8]

    Li,M.,Qiu,M.,Kong,W.,Zhu,L.,Ding,Y.:FusiongraphrepresentationofEEGfor emotionrecognition.Sensors23(3),1404(2023).https://doi.org/10.3390/s23031404

    work page doi:10.3390/s23031404 2023

  9. [9]

    Frontiers in Computational Neuroscience18, 1416494 (2024)

    Chen, W., Liao, Y., Dai, R., Dong, Y., Huang, L.: EEG-based emotion recognition using graph convolutional neural network with dual attention mechanism. Frontiers in Computational Neuroscience18, 1416494 (2024). https://doi.org/10.3389/fncom. 2024.1416494

    work page doi:10.3389/fncom 2024

  10. [10]

    In: Proceedings of the 5th International Confer- ence on Robotics and Control Engineering (RobCE), pp

    Ping, J., Xu, B., Wang, X., Zhang, W., Gao, Z., Song, A.: KAN-GCNN: EEG- based emotion recognition with a Kolmogorov-Arnold network-enhanced graph convolutional neural network. In: Proceedings of the 5th International Confer- ence on Robotics and Control Engineering (RobCE), pp. 44–49. ACM (2025). https://doi.org/10.1145/3733774.3735333

    work page doi:10.1145/3733774.3735333 2025

  11. [11]

    IEEE Access7, 93711–93722 (2019)

    Wang, Z., Tong, Y., Heng, X.: Phase-locking value based graph convolutional neural networks for emotion recognition. IEEE Access7, 93711–93722 (2019). https://doi.org/10.1109/ACCESS.2019.2927768

    work page doi:10.1109/access.2019.2927768 2019

  12. [12]

    Adversarial examples: Attacks and defenses for deep learning

    Hou, Y., et al.: GCNs-Net: A graph convolutional neural network approach for decoding time-resolved EEG motor imagery signals. IEEE Trans. Neural Networks 10 Wang et al. and Learning Systems35(6), 7312–7323 (2024). https://doi.org/10.1109/TNNLS. 2022.3202569

    work page doi:10.1109/tnnls 2024

  13. [13]

    IEEE Trans

    Song, T., Zheng, W., Song, P., Cui, Z.: EEG emotion recognition using dynamical graph convolutional neural networks. IEEE Trans. Affective Computing11(3), 532– 541 (2020). https://doi.org/10.1109/TAFFC.2018.2817622

    work page doi:10.1109/taffc.2018.2817622 2020

  14. [14]

    In: 2021 IEEE Conference on Com- putational Intelligence in Bioinformatics and Computational Biology (CIBCB), pp

    Ramakrishna, J.S., Sinha, N., Ramasangu, H.: Classification of human emotions using EEG-based causal connectivity patterns. In: 2021 IEEE Conference on Com- putational Intelligence in Bioinformatics and Computational Biology (CIBCB), pp. 1–8. IEEE (2021). https://doi.org/10.1109/CIBCB49929.2021.9562837

    work page doi:10.1109/cibcb49929.2021.9562837 2021

  15. [15]

    IEEE Trans

    Manomaisaowapak, P., Nartkulpat, A., Songsiri, J.: Granger causality inference in EEG source connectivity analysis: A state-space approach. IEEE Trans. Neural Networks and Learning Systems33(7), 3146–3156 (2022). https://doi.org/10.1109/ TNNLS.2021.3096642

    work page arXiv 2022

  16. [16]

    IEEE Trans

    Kong, W., Qiu, M., Li, M., Jin, X., Zhu, L.: Causal graph convolutional neural net- work for emotion recognition. IEEE Trans. Cognitive and Developmental Systems 15(4), 1686–1693 (2023). https://doi.org/10.1109/TCDS.2022.3175538

    work page doi:10.1109/tcds.2022.3175538 2023

  17. [17]

    https://doi.org/10.1109/ TNNLS.2023.3236635

    Ding, Y., Robinson, N., Tong, C., Zeng, Q., Guan, C.: LGGNet: learning from local-global-graphrepresentationsforbrain–computerinterface.IEEETrans.Neural Networks and Learning Systems35(7), 9773–9786 (2024). https://doi.org/10.1109/ TNNLS.2023.3236635

    work page arXiv 2024

  18. [18]

    In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp

    Xiao, M., Zhu, Z., Xie, K., Jiang, B.: MEEG and AT-DGNN: improving EEG emo- tion recognition with music introducing and graph-based learning. In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 4201–

    2024

  19. [19]

    https://doi.org/10.1109/BIBM62325.2024.10821806

    IEEE (2024). https://doi.org/10.1109/BIBM62325.2024.10821806

    work page doi:10.1109/bibm62325.2024.10821806 2024

  20. [20]

    Cognitive Neurodynamics20, 27 (2026)

    Gu, W., Peng, J., Ma, S., et al.: EEG emotion recognition based on hierarchical multi-scale graph neural networks. Cognitive Neurodynamics20, 27 (2026). https: //doi.org/10.1007/s11571-025-10396-x

    work page doi:10.1007/s11571-025-10396-x 2026

  21. [21]

    Entropy23(6), 679 (2021)

    Liang,X.S.:Normalizedmultivariatetimeseriescausalityanalysisandcausalgraph reconstruction. Entropy23(6), 679 (2021). https://doi.org/10.3390/e23060679

    work page doi:10.3390/e23060679 2021

  22. [22]

    Stips, A., Macias, D., Coughlan, C., Garcia-Gorriz, E., Liang, X.S.: On the causal structure between CO2 and global temperature. Sci. Rep.6(1), 2169 (2016). https: //doi.org/10.1038/srep21691

    work page doi:10.1038/srep21691 2016

  23. [23]

    Bai, C., Zhang, R., Bao, S., Liang, X.S., Guo, W.: Forecasting the tropical cyclone genesis over the Northwest Pacific through identifying the causal factors in cyclone– climate interactions. J. Atmos. Oceanic Technol.35(2), 247–259 (2018). https://doi. org/10.1175/JTECH-D-17-0109.1

    work page doi:10.1175/jtech-d-17-0109.1 2018

  24. [24]

    Remote Sens.9(6), 582 (2017)

    Zhang,H.,Qiu,Z.,Sun,D.,He,Y.:Seasonalandinterannualvariabilityofsatellite- derived chlorophyll-a (2000–2012) in the Bohai Sea, China. Remote Sens.9(6), 582 (2017). https://doi.org/10.3390/remotesensing.9.6.582

    work page doi:10.3390/remotesensing.9.6.582 2000

  25. [25]

    Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting

    Li, Y., Yu, R., Shahabi, C., Liu, Y.: Diffusion convolutional recurrent neural net- work: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  26. [26]

    Hierarchical Graph Representation Learning with Differentiable Pooling

    Ying, R., You, J., Morris, C., Ren, X., Hamilton, W.L., Leskovec, J.: Hierar- chical graph representation learning with differentiable pooling. arXiv preprint arXiv:1806.08804 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  27. [27]

    IEEE Trans

    Ding, Y., Tong, C., Zhang, S., Jiang, M., Li, Y., Lim, K.J., Guan, C.: EmT: a novel transformer for generalized cross-subject EEG emotion recognition. IEEE Trans. Neural Networks and Learning Systems36(6), 10381–10393 (2025). https: //doi.org/10.1109/TNNLS.2025.3552603

    work page doi:10.1109/tnnls.2025.3552603 2025

  28. [28]

    In: 2020 IEEE International Conference on Sys- tems, Man, and Cybernetics (SMC), pp

    Ingolfsson, T.M., Hersche, M., Wang, X., Kobayashi, N., Cavigelli, L., Benini, L.: EEG-TCNet: an accurate temporal convolutional network for embedded motor- GL-LFGNN 11 imagery brain–machine interfaces. In: 2020 IEEE International Conference on Sys- tems, Man, and Cybernetics (SMC), pp. 2958–2965. IEEE (2020). https://doi.org/ 10.1109/SMC42975.2020.9283028

    work page doi:10.1109/smc42975.2020.9283028 2020

  29. [29]

    IEEE Trans

    Altaheri, H., Muhammad, G., Alsulaiman, M.: Physics-informed attention tempo- ral convolutional network for EEG-based motor imagery classification. IEEE Trans. Industrial Informatics19(2), 2249–2258 (2023). https://doi.org/10.1109/TII.2022. 3197419

    work page doi:10.1109/tii.2022 2023