Scholarlink

<p>EEG-based emotion recognition has commonly been studied under classification settings, where emotional states are mapped to a set of discrete categories. In practice, however, emotional experiences are continuous and evolve over time. From this perspective, regression-based approaches are better suited for emotion modeling, as they allow continuous estimation of emotional dimensions such as valence, arousal, dominance, and liking, thereby providing a more detailed characterization of affective dynamics. Nevertheless, accurate EEG-based emotion regression remains difficult, mainly due to the low signal-to-noise ratio of EEG recordings, pronounced inter-subject variability, and the complex spatiotemporal organization of brain activity. To address these challenges, the Parallel Feature-fusion Graph Convolutional Network (PFGCN) is proposed. The model utilizes parallel graph convolutional branches with varied dilation rates to extract heterogeneous spatial dependencies at multiple granularities, providing the resolution necessary to track nuanced emotional fluctuations. These multi-scale features are then refined by a Feature-wise Linear Modulation (FiLM) layer. By adaptively scaling and shifting feature maps, this mechanism recalibrates node importance to prioritize discriminative neural signatures over noise in non-stationary EEG signals, ensuring the representational precision required for robust continuous emotion regression. In addition, temporal dependencies in EEG sequences are modeled using a long short-term memory (LSTM) network, enabling the capture of temporal variations in emotional states. Experimental results on the DEAP and MAHNOB-HCI datasets, evaluated using a 10-fold cross-validation protocol, demonstrate that PFGCN yields lower prediction errors across all four emotional dimensions compared to several representative baseline approaches. These findings suggest that PFGCN constitutes a viable framework for continuous EEG-based emotion recognition and contributes to the analysis of spatiotemporal patterns associated with emotional processing.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive spatial-temporal graph modeling for continuous emotion prediction based on EEG

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Boxun Meng,
  • Zhengnan Zhang,
  • Jiangwen Lu,
  • Yunyuan Gao,
  • Yanhua Qin

摘要

EEG-based emotion recognition has commonly been studied under classification settings, where emotional states are mapped to a set of discrete categories. In practice, however, emotional experiences are continuous and evolve over time. From this perspective, regression-based approaches are better suited for emotion modeling, as they allow continuous estimation of emotional dimensions such as valence, arousal, dominance, and liking, thereby providing a more detailed characterization of affective dynamics. Nevertheless, accurate EEG-based emotion regression remains difficult, mainly due to the low signal-to-noise ratio of EEG recordings, pronounced inter-subject variability, and the complex spatiotemporal organization of brain activity. To address these challenges, the Parallel Feature-fusion Graph Convolutional Network (PFGCN) is proposed. The model utilizes parallel graph convolutional branches with varied dilation rates to extract heterogeneous spatial dependencies at multiple granularities, providing the resolution necessary to track nuanced emotional fluctuations. These multi-scale features are then refined by a Feature-wise Linear Modulation (FiLM) layer. By adaptively scaling and shifting feature maps, this mechanism recalibrates node importance to prioritize discriminative neural signatures over noise in non-stationary EEG signals, ensuring the representational precision required for robust continuous emotion regression. In addition, temporal dependencies in EEG sequences are modeled using a long short-term memory (LSTM) network, enabling the capture of temporal variations in emotional states. Experimental results on the DEAP and MAHNOB-HCI datasets, evaluated using a 10-fold cross-validation protocol, demonstrate that PFGCN yields lower prediction errors across all four emotional dimensions compared to several representative baseline approaches. These findings suggest that PFGCN constitutes a viable framework for continuous EEG-based emotion recognition and contributes to the analysis of spatiotemporal patterns associated with emotional processing.