A Cross-Modality Graph Attention Network for Robust Multimodal Emotion Recognition with Missing-Modality Handling

Abstract

Multimodal emotion recognition benefits from combining physiological and behavioral signals, but real-world systems often face missing, noisy, or asynchronous modalities. This paper presents a rewritten IEEE-style research article on a crossmodality graph attention network for robust emotion recognition using EEG, ECG, GSR, and facial features. The proposed MAGFusionNet framework constructs modality-specific graphs, learns intra-modal representations through graph attention layers, and applies reliability-gated fusion to reduce performance degradation when one or more modalities are missing. A residual recovery branch estimates missing-modality context from available signals. Simulation-based evaluation inspired by DEAP, DREAMER, and MAHNOB-HCI input structures demonstrates improved accuracy and F1-score compared with early fusion, late fusion, CNN-LSTM, and non-gated graph fusion baselines. The paper provides original wording, editable workflow diagrams, mathematical equations, comparative tables, and consistent results suitable for student publication