PaperlayerView exact revision on arXiv

Source-linked AI summary

CDBG: Causally Motivated Dual-Invariance Learning against Topological and Predictive Shifts in EEG Workload Recognition

Yuzhe Zhang, Wenmin Zhou, Chengxi Xie, Kai He, Jihong Wang, Huan Liu, Man Yao, Daoqiang Zhang

arXiv:2609.30831v1cs.HC

TL;DR

Inter-subject variability makes EEG workload recognition on unseen subjects difficult, with functional graphs facing shifts in both rationale topology and prediction mappings. CDBG combines rationale selection with topology alignment and subject-wise IRM. Under strict LOSO evaluation, it improves Macro-F1 over state-of-the-art baselines and yields interpretable functional rationales.

  • Problem

    Subject variability makes EEG workload recognition on unseen subjects difficult, with shifts in both class-conditional functional topology and the representation-to-label mapping.

    Source: Introduction · p. 1

  • Method

    CDBG uses stochastic edge masking to select workload-predictive graph rationales, workload-conditional Laplacian spectral alignment, and subject-wise IRM.

    Source: Introduction · p. 2

  • Results

    Macro-F1 scores improve by 2.48%, 3.73%, and 4.23% on SELF, STEW, and EEGMAT, respectively, over state-of-the-art baselines under LOSO.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    Under strict LOSO evaluation, CDBG achieves state-of-the-art performance while yielding neurophysiologically interpretable functional rationales.

    Source: Introduction · p. 2, Conclusion · p. 7

    The target spectral distribution is assumed to be a convex mixture of source distributions, covering interpolation within observed variation rather than arbitrary extrapolation.

    Source: Target Spectral Discrepancy under… · p. 4

Abstract

from arXiv · show

Generalizing Electroencephalography (EEG)-based mental workload recognition to unseen subjects remains a formidable challenge due to severe inter-subject variability. While functional brain graphs effectively model distributed cognitive dynamics, their inherent subject-specificity induces two coupled distribution shifts: a class-conditional topological shift in the underlying functional connectivity, and a predictive mechanism shift in the learned representation-to-label mapping. Motivated by the subject-induced distribution shifts, we propose CDBG, a Causally motivated Dual-invariance learning framework for Brain Graphs. CDBG disentangles and mitigates these shifts via a two-stage rationale learning pipeline. First, it employs stochastic edge masking to extract sparse, workload-predictive graph rationales, regularized by workload-conditional Laplacian spectral alignment to enforce topological invariance across subjects. Second, it applies subject-wise Invariant Risk Minimization (IRM) to the graph representations, ensuring environment-wise risk stationarity. Extensive experiments on a self-built air traffic controller EEG cognitive workload dataset and multiple public datasets under a strict leave-one-subject-out protocol demonstrate that CDBG significantly outperforms state-of-the-art cross-subject and graph-based baselines, improving the Macro-F1 score by up to 4.23%, while simultaneously providing neurophysiologically interpretable functional rationales.

Introduction

Cross-subject workload recognition is complicated by subject-specific functional connectivity, which produces coupled shifts in rationale topology and prediction. CDBG addresses both shifts, and its LOSO evaluation reports higher Macro-F1 scores than state-of-the-art baselines.

Introduction

  • Subject variability creates class-conditional topology shifts during rationale selection and predictive-mechanism shifts during classification, while existing rationale selection can exploit subject-specific connectivity.Even with stable topology, subject-specific node-feature baselines can shift the representation-to-prediction mapping.Source: Introduction · pp. 1–2
  • CDBG treats each training subject as an environment, using workload-conditional Laplacian spectral alignment and subject-wise IRM to target topology and prediction shifts.Joint optimization shapes subgraphs for workload predictiveness, sparsity, and cross-subject stability.Source: Introduction · p. 2
  • Macro-F1 scores improve by 2.48%, 3.73%, and 4.23% on SELF, STEW, and EEGMAT, respectively, over state-of-the-art baselines under LOSO.The evaluation uses one self-built and two public EEG datasets.Source: Introduction · p. 2
  • CDBG extracts predictive connectivity through stochastic edge masking and combines spectral alignment with subject-wise IRM regularization.The contributions also include a theoretical bound linking source spectral alignment to unseen-subject discrepancy and a decomposition of the subject-wise IRM penalty.Source: Introduction · p. 2
  • Strict LOSO experiments and graph-level analyses support the framework’s effectiveness and interpretability of its functional rationales.The paper identifies workload recognition as important for adaptive human-machine systems and describes functional graphs as models of channel-wise activity and connectivity.Source: Introduction · pp. 1–2

Related Work

Prior EEG workload methods progressed from hand-crafted and deep signal-processing approaches to graph models of channel interactions. Existing cross-subject and invariant-learning approaches leave workload-conditional functional topology insufficiently addressed.

Related Work

  • EEG workload pipelines progressed from hand-crafted features to deep spectral-temporal models and then graph models representing EEG channels and their interactions.The cited approaches include adaptive adjacency and biological topology in early EEG GNNs, followed by attention- and structure-learning variants.Source: Related Work · p. 2
  • Cross-subject methods align representations through adaptation, adversarial inference, mixture-of-experts, contrastive learning, and self-supervised pretraining, but do not directly regularize functional connectivity structure.Domain adaptation requires target-subject calibration data, unlike subject-independent generalization from source subjects alone.Source: Related Work · p. 2
  • Invariant rationale methods such as CIGA [Chen et al. 2022] and GSAT [Miao, Liu, and Li 2022] extract predictive subgraphs, but are not specialized for subject-indexed brain graphs or coupled topology and node-attribute variation.CDBG treats each subject as an environment and targets both topological and predictive-mechanism shifts.Source: Related Work · p. 2

Method

CDBG treats each training subject as an environment and learns a graph rationale under topological and predictive invariance constraints. Its pipeline aligns workload-conditional graph spectra and uses subject-wise IRM to promote stable prediction across environments.

Method

  • Given a functional graph, CDBG’s learnable selector extracts a sparse rationale that a graph encoder and classifier map to a prediction.Each training subject is treated as a distinct environment for the framework’s invariance constraints.Source: Method · p. 2, Figure 2 · p. 3

Problem Setup and Dual-Invariance Goals

  • CDBG targets stable class-conditional rationale distributions and stable representation-to-label relationships across training-subject environments.The two goals correspond to topological invariance at selection and predictive invariance at classification.Source: Method · p. 2, Problem Setup and Dual-Invariance Goals · p. 3

Mitigating Class-Conditional Topological Shift

  • A stochastic edge selector extracts a predictive subgraph, while workload-conditional Laplacian spectral alignment mitigates class-conditional topological shift across subjects.The method aligns spectral descriptors for samples in the same workload class using MMD.Source: Mitigating Class-Conditional… · pp. 3–4

Mitigating Cross-Subject Predictive Shift

  • Because spectral alignment overlooks predictive variation in subject-specific node features, CDBG adds subject-wise IRMv1 to promote a shared, environment-stationary prediction rule.IRMv1 penalizes non-stationarity along a shared logit-scaling direction as a practical surrogate for predictive invariance.Source: Mitigating Cross-Subject Predictive… · p. 4

Joint Optimization and Inference

  • CDBG jointly optimizes its rationale-learning and classification objectives, requiring subject indices only during training and using a deterministic inference path without test-time adaptation.The inference path maps an unseen subject’s graph through the rationale selector, graph encoder, and classifier.Source: Joint Optimization and Inference · p. 4

Theoretical Analysis of Dual-Invariance

The theoretical analysis bounds unseen-subject spectral discrepancy under a source-coverage assumption and decomposes the subject-wise IRM penalty into average stationarity and cross-subject gradient variance.

Theoretical Analysis of Dual-Invariance

Target Spectral Discrepancy under Source Coverage

  • Under the source-coverage assumption, source spectral loss bounds the target-versus-source-mixture discrepancy for class-specific spectral statistics.The target spectral distribution is assumed to be a convex mixture of source distributions, representing interpolation within source-observed variation.Source: Target Spectral Discrepancy under… · p. 4

Decomposition of the Subject-Wise IRM Penalty

Experimental Settings

The experiments compare CDBG with twelve baselines on three EEG workload datasets using strict leave-one-subject-out evaluation, with shared preprocessing and validation-based model selection.

Experimental Settings

  • The benchmark setup specifies comparison methods, evaluation protocol, datasets, and implementation configurations.Extended dataset statistics, baseline descriptions, dataset-specific settings, and training parameters are reported in supplementary material.Source: Experimental Settings · p. 5

Comparison Methods

  • Twelve baselines span graph-based, workload-specific, and general EEG models, covering convolutional, attention-based, and graph-based representation learning.The compared methods include four models in each category.Source: Comparison Methods · p. 5

Evaluation Protocol

  • Strict LOSO holds one subject out for testing and uses the remaining subjects for training and model selection; results are averaged across folds.Accuracy and Macro-F1 are reported, with Macro-F1 as the primary metric.Source: Evaluation Protocol · p. 5

Datasets

  • SELF, EEGMAT, and STEW vary in channel count, subject count, sampling rate, and workload conditions.SELF comprises 59-channel recordings from eight air traffic controllers; EEGMAT has 19 channels and 36 subjects, while STEW has 14 channels and 48 subjects.Source: Datasets · p. 5

Preprocessing and Implementation Details

  • All methods use the same processed inputs, LOSO partitions, and evaluation protocol, with validation Macro-F1 guiding checkpoint retention and hyperparameter selection.Signals are bandpass- and notch-filtered, segmented into non-overlapping one-second windows, and channel-wise z-score normalized.Source: Preprocessing and Implementation Details · p. 5

Experiments Results and Analysis

CDBG outperforms the strongest baselines across three datasets, and component analyses support the complementary roles of its regularizers. Connectivity and representation analyses show workload-sensitive patterns and progressively clearer workload separation under the full model.

Comparison Experiments

  • CDBG ranks first in ACC and Macro-F1 across three datasets, exceeding strongest baselines by 2.48, 4.23, and 3.73 Macro-F1 points; over 80% of comparisons were significant.The gains are reported for SELF, EEGMAT, and STEW, respectively; significance tests use subject-level ACC and Macro-F1.Source: Comparison Experiments · p. 6

Ablation Study

  • Removing spectral alignment, IRM, or adaptive masking reduces performance, with IRM removal causing the largest Macro-F1 drops on SELF and EEGMAT.The ablations indicate that spectral alignment helps preserve stable graph structures, IRM supports invariant workload representations, and masking retains informative edges while suppressing irrelevant ones.Source: Ablation Study · p. 6

Hyperparameter Analysis

  • On SELF, best performance uses λsp = 0.8, λIRM = 0.4, and λmask = 1.0; tested-range performance has single-peak trends.Moderate spectral alignment balances topology alignment with discriminative structure, while overemphasizing the IRM gradient penalty degrades performance; λmask = 1.0 controls density and suppresses redundant edges.Source: Hyperparameter Analysis · pp. 6–7, Figure 3 · p. 7

Workload-Dependent Connectivity Interpretation

  • On EEGMAT, low-workload graphs emphasize frontal, central, and parieto-occipital connections, while high workload highlights Cz, Pz, P4, and O2 with stronger cross-regional integration.Figure 4 encodes connection strength by edge width and incident-connection contribution by node size.Source: Workload-Dependent Connectivity… · p. 7, Figure 4 · p. 7

Cross-Subject Representation Analysis

  • On held-out STEW samples, masking removes irrelevant connectivity, spectral alignment brings same-class samples across subjects closer, and IRM further separates workloads, yielding the clearest class boundaries and weakest subject clustering.The comparison progresses from raw graph features through masking and spectral alignment to full CDBG with IRM.Source: Cross-Subject Representation Analysis · p. 7, Figure 5 · p. 7

Conclusion

The conclusion frames cross-subject workload recognition as involving topological and predictive shifts and presents CDBG as a dual-invariance approach. It reports state-of-the-art LOSO performance with interpretable rationales and identifies other EEG cognitive tasks and unsupervised test-time adaptation as future work.

Conclusion

  • CDBG jointly uses stochastic graph rationale selection, workload-conditional Laplacian spectral alignment, and subject-wise IRM to address topological and predictive mechanism shifts.The framework is described as separating task-intrinsic neural signatures from subject-specific variations.Source: Conclusion · p. 7
  • Under strict leave-one-subject-out protocols, CDBG achieves state-of-the-art performance while yielding neurophysiologically interpretable functional rationales.The supplied conclusion does not specify numerical results.Source: Conclusion · pp. 7–8
  • Future work will extend CDBG to other EEG-based cognitive tasks and explore unsupervised test-time adaptation for real-world deployment.Source: Conclusion · p. 8
Loading 2609.30831v1…