🤖 AI Summary
To address the challenges of insufficient high-quality emotion labels in naturalistic scenarios and domain bias in physiological signal representations, this paper proposes a weakly supervised deep clustering framework. Our method introduces the novel “deep seeded clustering” paradigm, which jointly optimizes unsupervised representation learning via autoencoders and dynamic fuzzy c-means (FCM) clustering, integrated with multi-source physiological signal preprocessing and feature disentanglement to enable end-to-end fine-grained emotion classification without manual annotations. Evaluated on three benchmark datasets—WESAD, Stress-Predict, and CEAP360-VR—the framework achieves accuracies of 80.7%, 64.2%, and 61.0%, respectively. These results demonstrate substantial improvements in generalization and robustness for real-world, label-scarce environments, advancing weakly supervised affective computing with physiological signals.
📝 Abstract
According to the circumplex model of affect, an emotional response could characterized by a level of pleasure (valence) and intensity (arousal). As it reflects on the autonomic nervous system (ANS) activity, modern wearable wristbands can record non-invasively and during our everyday lives peripheral end-points of this response. While emotion recognition from physiological signals is usually achieved using supervised machine learning algorithms that require ground truth labels for training, collecting it is cumbersome and particularly unfeasible in naturalistic settings, and extracting meaningful insights from these signals requires domain knowledge and might be prone to bias. Here, we propose and test a deep-seeded clustering algorithm that automatically extracts and classifies features from those physiological signals with minimal supervision - combining an autoencoder (AE) for unsupervised feature representation and c-means clustering for fine-grained classification. We also show that the model obtains good performance results across three different datasets frequently used in affective computing studies (accuracies of 80.7% on WESAD, 64.2% on Stress-Predict and 61.0% on CEAP360-VR).
