![[2510.03027] Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2510.03027] Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling
Summary
This article presents a novel lightweight transformer model for EEG classification, utilizing a balanced signed graph algorithm to enhance performance while reducing parameters.
Why It Matters
The research addresses the challenge of classifying EEG signals, particularly for epilepsy detection, by introducing an efficient model that balances performance and interpretability. This advancement could lead to better diagnostic tools in neurology, making EEG analysis more accessible and effective.
Key Takeaways
- Introduces a lightweight transformer model for EEG classification.
- Utilizes balanced signed graphs to model EEG signal anti-correlations.
- Achieves competitive classification performance with fewer parameters.
- Implements an efficient low-pass filter using Lanczos approximation.
- Demonstrates potential for improved diagnostic tools in neurology.
Computer Science > Machine Learning arXiv:2510.03027 (cs) [Submitted on 3 Oct 2025 (v1), last revised 14 Feb 2026 (this version, v3)] Title:Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling Authors:Junyi Yao, Parham Eftekhar, Gene Cheung, Xujin Chris Liu, Yao Wang, Wei Hu View a PDF of the paper titled Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling, by Junyi Yao and 5 other authors View PDF HTML (experimental) Abstract:Samples of brain signals collected by EEG sensors have inherent anti-correlations that are well modeled by negative edges in a finite graph. To differentiate epilepsy patients from healthy subjects using collected EEG signals, we build lightweight and interpretable transformer-like neural nets by unrolling a spectral denoising algorithm for signals on a balanced signed graph -- graph with no cycles of odd number of negative edges. A balanced signed graph has well-defined frequencies that map to a corresponding positive graph via similarity transform of the graph Laplacian matrices. We implement an ideal low-pass filter efficiently on the mapped positive graph via Lanczos approximation, where the optimal cutoff frequency is learned from data. Given that two balanced signed graph denoisers learn posterior probabilities of two different signal classes during training, we evaluate their reconstruction errors for binary classification of EEG signals. Experiments show that our...
