Boosting EEG emotion recognition: A multi-dataset study of MIL-augmented XGBoost classifiers

Türkçe

Background: EEG-based emotion recognition has attracted growing interest in affective computing due to its high temporal resolution in capturing neural responses. However, EEG signals are inherently noisy and non-stationary, and many temporal segments within a trial may not reflect the underlying emotional state. Conventional pipelines assume all segments contribute equally to the trial label, reducing robustness when irrelevant or transitional segments are present. New Method: This study proposes a Multiple Instance Learning (MIL) framework integrated with eXtreme Gradient Boosting (XGBoost) for EEG emotion recognition. Each trial is treated as a bag of temporal segment instances, with the emotional label assigned at the bag level. During training, bag labels are propagated to instances for instance-level classification with XGBoost. During inference, instance-level probabilities are aggregated via mean pooling to yield bag-level predictions, enabling the model to emphasize informative signal fragments while suppressing noise. Results: The framework is evaluated on two benchmark datasets, DEAP and VREMO, across multiple feature representations including preprocessed time-series, time-domain, frequency-domain, time-frequency, decomposition-based, and spatial features. Comparison with Existing Methods: MIL-XGBoost consistently outperforms standard XGBoost across all feature domains and classification tasks, achieving up to 94–95% accuracy for binary emotion classification on DEAP and competitive results on VREMO, without requiring synthetic augmentation or deep neural architectures. Conclusion: Integrating MIL with gradient-boosted tree models offers an efficient and robust alternative for weakly supervised EEG emotion recognition, demonstrating consistent gains under realistic labeling assumptions.