کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Emotion Recognition Using Effective Connectivity and Fully Complex-Valued Magnetic Graph Convolution Neural Network

نویسندگان :

Armin Pishehvar¹ Eghbal Mansoori² Abbas Mehrbaniyan³ Reza Tahmasebi⁴

1- دانشگاه شیراز 2- دانشگاه شیراز 3- دانشگاه شیراز 4- دانشگاه شیراز

کلمات کلیدی :

emotion recognition،Electroencephalogram،graph convolutional neural networks،effective connectivity

چکیده :

Emotion recognition plays a vital role in our lives, from fostering deeper social connections and improving communication to enhancing human-computer interaction and personalizing healthcare. However, accurately deciphering these internal states, especially through physiological signals, presents a significant challenge. Among various methods, emotion recognition using electroencephalography (EEG) has been a persistent area of research, though it faces unique complexities. While much work in EEG-based emotion recognition emphasizes the classification of broad categories like positive versus negative emotions, the exploration of multi-class emotion recognition encompassing a wider spectrum, such as nine distinct emotional states, remains largely underexplored. To address this critical gap, we introduce FCMagnet, a novel fully complex-valued magnetic graph convolutional network, uniquely designed for nine-class EEG-based emotion recognition. Unlike traditional real-valued graph neural networks, FCMagnet captures directed effective brain connectivity through multivariate autoregressive (MVAR) modeling and partial directed coherence (PDC), encoding these as Hermitian Laplacians to preserve both magnitude and phase information. This approach, leveraging complex spectral filtering and complex-domain activation, learns rich representations of emotional brain states. Evaluated on the large FACED dataset, FCMagnet achieved 31.2 ± 3.6% accuracy, substantially outperforming classical real-valued GNNs and matching state-of-the-art spatial models while remaining more compact and interpretable. Our results clearly show that fully complex-valued spectral graph filtering provides a powerful and interpretable framework for advancing fine-grained emotion recognition.