A logic-based knowledge-driven bidirectional multi-attention GRU framework for fear level classification in humans

Abstract The importance of fear level detection in affective computing using physiological signals has not been extensively explored. Accurately detecting fear responses facilitates real-time emotional monitoring and improves therapeutic outcomes. The logic-based learning (LBL) method integrates data-driven learning and symbolic reasoning, using logical predicates to model knowledge and develop interpretable models. By combining domain expertise with computational intelligence, the proposed framework extracts key features such as Frontal Alpha Asymmetry (FAA) from electroencephalograms (EEG) and EMO signals, and peak values from Galvanic Skin Responses (GSR) signals. These extracted features are used to determine feature-specific thresholds and logical rules. To achieve high classification performance, logic rules are defined using an Inductive Logic Programming (ILP) model. To evaluate performance metrics, data obtained from ILP are analyzed. To enhance interpretability and classification, logic rules derived from the ILP model are incorporated into deep learning models. This study uses several deep learning models for detecting fear, including Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Gated Recurrent Units (GRU), and Bidirectional Multi-Attention GRU (Bi-MA-GRU). This study examines the Dataset for Emotion Analysis using Physiological Signals (DEAP) and Multimodal Database for Decoding Affective Physiological Responses (DECAF) to facilitate comparisons. Using multimodal physiological data, deep learning models, and LBL techniques, the proposed technique shows that fear levels can be detected in a scalable manner through the integration of multimodal physiological data, deep learning models, and LBL techniques. A knowledge-based approach and computational intelligence enhance methodologies for understanding and managing fear-related conditions. Thus, the formulated technique achieved better performance metrics using a Bidirectional Multi-Attention GRU model that had a 96.67% accuracy with DEAP dataset and music-based data on DECAF versus 86.67% accuracy with movie-based data.