Interpretable Multi-Task PINN for Emotion Recognition and EDA Prediction

The WESAD dataset is a widely used multimodal benchmark for wearable stress and affect detection. In this work, we present a processed and feature-engineered version of the WESAD dataset, specifically curated for emotion recognition and Electrodermal Activity (EDA) prediction tasks. The dataset includes synchronized and standardized input features extracted from raw physiological signals, including EDA, respiration, and body temperature, along with emotion labels mapped to discrete affective states.In addition to the original signals, we provide a suite of derived features such as statistical signal descriptors, frequency-domain representations, and smoothed temporal segments. These features are intended to support research on interpretable and multi-task deep learning frameworks, such as Physics-Informed Neural Networks (PINNs) for emotion-aware modeling.The dataset is cleaned, annotated, and formatted as CSV files, making it readily accessible for training, evaluation, and benchmarking of multimodal affective computing models. All preprocessing steps and label mappings are documented to ensure reproducibility.