[2026 IEEE TCDS Special Issue] Distance-Structured Lag-Sparse Graph Convolution with a Lag-Shared Adjacency Residual for Imagined Speech EEG Decoding

Seung Won Kim, Dae Hyeon Kim, Young-Seok Choi^*

Department of Electronics and Communications Engineering, Kwangwoon University, Seoul, South Korea

---

📢 News

• [Aug. 2026] 📄 Our paper "Distance-Structured Lag-Sparse Graph Convolution with a Lag-Shared Adjacency Residual for Imagined Speech EEG Decoding" has been submitted to the IEEE Transactions on Cognitive and Developmental Systems (TCDS) Special Issue. The code will be officially released upon acceptance.

---

📝 Abstract

Electroencephalographic imagined-speech decoding must represent electrode interactions over time from a limited number of subject-specific trials. Unified spatial–temporal window graphs provide direct lagged electrode interactions, but a τ-frame residual adjacency scales as τ²N² for N electrodes.

We propose a distance-structured lag-sparse graph convolutional network (DLS-GCN). Each branch represents temporal context by selected shifts and an N × N spatial operator instead of a τN × τN window graph. Exact-distance supports through two hops define distance-specific normalization and the support of adaptive residuals. One masked residual per distance component is shared across all temporal lags of a branch. This constraint permits exact factorization of the lag dimension; because the distance components are also summed before nonlinear processing, they can be folded into one effective graph product per branch without changing the output. A parameter-free temporal recalibration module and a depthwise multiscale temporal module form the remaining network.

Under protocol-matched subject-dependent evaluation, DLS-GCN obtains 0.8433 ± 0.0459 accuracy on Track 3 of the 2020 International BCI Competition and 0.4108 ± 0.0260 on the Thinking Out Loud dataset. The trained model has 430,853 parameters and 166.88 M MACs in the component-wise implementation; exact graph folding reduces evaluation to 160.39 M MACs. Controlled analyses support the distance-structured prior and lag-shared residual as an effective accuracy–complexity trade-off.

---

📊 Experimental Results

Protocol-matched subject-dependent evaluation.

Dataset	Accuracy
2020 International BCI Competition (Track 3)	0.8433 ± 0.0459
Thinking Out Loud	0.4108 ± 0.0260

Complexity

Metric	Value
Parameters	430,853
MACs (component-wise implementation)	166.88 M
MACs (exact graph folding, evaluation)	160.39 M

Key Findings:

1. Lag factorization. Sharing one masked residual per distance component across all temporal lags of a branch permits exact factorization of the lag dimension, replacing the τN × τN window graph with an N × N spatial operator.
2. Exact graph folding. Because the distance components are summed before nonlinear processing, they fold into a single effective graph product per branch without changing the output, reducing evaluation cost from 166.88 M to 160.39 M MACs.
3. Accuracy–complexity trade-off. Controlled analyses support the distance-structured prior and the lag-shared residual as an effective trade-off between accuracy and complexity.

---

📦 Repository Structure

This repository contains the minimal, self-contained code required to reproduce training and evaluation on Track 3 of the 2020 International BCI Competition.

main.py             Full protocol runner (15 subjects x 10 seeds) and result aggregation
train_eval.py       One run (single subject, single seed): preprocessing, training, evaluation
preprocessing.py    .mat loading, signal preprocessing, seed-wise data splitting
model.py            Model definition

File	Role
model.py	Model definition. EEG channels are treated as graph nodes and inter-channel WPLI phase synchronization as edges. Contains the WPLI graph builder, exact-hop k-adjacency, the dual-branch aggregation with lag-shared masked residuals, the parameter-free temporal recalibration module, and the depthwise multi-scale temporal module. Input (B, 64, 795) → class logits (B, 5).
preprocessing.py	Loads the per-subject .mat files, concatenates them, and re-splits them per seed into 60/10/10 trials per class. Applies common average reference, a 60 Hz low-pass filter, and channel-wise z-scoring whose statistics are fitted on the training split only.
train_eval.py	One complete run: seeding → preprocessing → graph initialization → training (AdamW, polynomial LR decay, mixed precision, L_cls + λ·L_sparse) → best-validation checkpoint restoration → test evaluation, with learning curves, confusion matrix, and classification report written to disk.
main.py	Runs the full protocol over 15 subjects × 10 seeds and aggregates per-subject and overall accuracy.

Data

Place Data_Sample{1..15}.mat (containing epo_train / epo_validation / epo_test) in the following directories:

Training set/     Data_Sample1.mat ... Data_Sample15.mat
Validation set/   Data_Sample1.mat ... Data_Sample15.mat
Test set/         Data_Sample1.mat ... Data_Sample15.mat

Quick Start

python main.py

Results are written to results/proposed_final/subject_XX/seed_YYYY/, with ALL_SUBJECTS_SUMMARY.txt and RESULT_{mean}_{std}.txt at the top level.

Key Hyperparameters

Argument	Value	Description
NUM_EPOCHS	200	Training epochs
BATCH_SIZE	64	Mini-batch size
FS	256	Sampling rate (Hz)
LAMBDA_SPARSE	1e-4	L1 weight on the masked residual adjacency
SEEDS	10 seeds	Each seed defines a different 60/10/10 per-class split
SUBJECT_IDS	1–15	Subject-dependent evaluation

Optimizer: AdamW, learning rate 1e-3 → 1e-6 (polynomial decay, power 2), weight decay 1e-2, cross-entropy with label smoothing 0.01.

---

📄 License

This project is released under the MIT License. See LICENSE for details.

---

Citation

The paper is currently under review at the IEEE Transactions on Cognitive and Developmental Systems (TCDS) Special Issue. A citation entry will be added upon acceptance.