A lightweight cross-session EEG emotion recognition network based on high- and low-frequency differential weighting

Non-invasive electroencephalography (EEG) holds promising potential for brain-computer interfaces and neural engineering. However, cross-session distribution shift and the computational overhead of existing models remain major bottlenecks for deployment on wearable devices. To address these challenges, this paper first verifies the effectiveness of full-band features and reveals the complementary patterns between high- and low-frequency features through progressive frequency-band ablation experiments. Building on these findings, we propose FreqW-LiteNet, a lightweight network with frequency-band differentiated weighting. Avoiding complex modules such as multi-source domain adaptation or adversarial training, the model adopts a single-backbone fully connected architecture, a two-stage training strategy, and high-intensity regularization to achieve efficient training on standard CPUs. Systematic validation is conducted on three public datasets (SEED, SEED-IV, and SEED-V) under cross-session protocols using a transductive evaluation setting. All experiments are strictly restricted to these SEED-family datasets. Experimental results show that, on the SEED and SEED-IV datasets, FreqW-LiteNet achieves competitive recognition accuracy compared with state-of-the-art domain adaptation methods and general deep learning models. Notably, the model is extremely compact, with approximately 23,000 parameters, considerably reducing computational overhead. The research demonstrates that the synergy among frequency-band differentiated weighting, two-stage training, and high-intensity regularization enables the lightweight architecture to achieve stable cross-session EEG emotion recognition without complex domain adaptation modules, offering a viable solution that balances performance and efficiency for EEG decoding in resource-constrained scenarios.

Authors

Institutions

Publication Details

Journal
Frontiers in Neuroscience
Published
2026-11-19
DOI
https://doi.org/10.3389/fnins.2026.1992372
Primary Topic
Emotion and Mood Recognition
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A lightweight cross-session EEG emotion recognition network based on high- and low-frequency differential weighting

Rui Fan, ZhengXiang Wan, Houru Yan, Junfeng Hu
Frontiers in Neuroscience
Emotion and Mood Recognition
article

A lightweight cross-session EEG emotion recognition network based on high- and low-frequency differential weighting

Rui Fan, ZhengXiang Wan, Houru Yan, Junfeng Hu
article en

Abstract

Non-invasive electroencephalography (EEG) holds promising potential for brain-computer interfaces and neural engineering. However, cross-session distribution shift and the computational overhead of existing models remain major bottlenecks for deployment on wearable devices. To address these challenges, this paper first verifies the effectiveness of full-band features and reveals the complementary patterns between high- and low-frequency features through progressive frequency-band ablation experiments. Building on these findings, we propose FreqW-LiteNet, a lightweight network with frequency-band differentiated weighting. Avoiding complex modules such as multi-source domain adaptation or adversarial training, the model adopts a single-backbone fully connected architecture, a two-stage training strategy, and high-intensity regularization to achieve efficient training on standard CPUs. Systematic validation is conducted on three public datasets (SEED, SEED-IV, and SEED-V) under cross-session protocols using a transductive evaluation setting. All experiments are strictly restricted to these SEED-family datasets. Experimental results show that, on the SEED and SEED-IV datasets, FreqW-LiteNet achieves competitive recognition accuracy compared with state-of-the-art domain adaptation methods and general deep learning models. Notably, the model is extremely compact, with approximately 23,000 parameters, considerably reducing computational overhead. The research demonstrates that the synergy among frequency-band differentiated weighting, two-stage training, and high-intensity regularization enables the lightweight architecture to achieve stable cross-session EEG emotion recognition without complex domain adaptation modules, offering a viable solution that balances performance and efficiency for EEG decoding in resource-constrained scenarios.

Frontiers in NeuroscienceVol. 20
Xuzhou Medical College (CN)
Emotion and Mood Recognition
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon