M-GGF: A Multidimensional Graph with Adaptive Graph Gated Fusion for Multimodal Emotion Recognition

Emotion recognition, a central task in affective computing, is pursued with the objective of deriving a deeper and more robust interpretation of human emotional states through the integration of multimodal data. Although existing multimodal emotion recognition approaches have made notable progress, they still face substantial challenges in modeling complex inter-modal interactions, long-range semantic dependencies, and effective information fusion. To address these issues, this study proposes a Multimodal Emotion Recognition Model based on Multidimensional Graphs and Adaptive Graph-Gated Fusion (M-GGF). Within M-GGF, a Multidimensional Graph Construction (MGC) module is developed to build a unified heterogeneous graph that jointly models temporal dependencies, cross-modal interactions, and long-range semantic relations. In addition, an Adaptive Graph-Gated Residual Fusion (Adaptive GGRF) mechanism is introduced to achieve fine-grained integration of local and global semantics. Experimental results on the IEMOCAP benchmark confirm the superior performance of the proposed M-GGF model. Specifically, it achieves accuracies of 68.53% and 85.28% alongside weighted F1 scores of 68.76% and 84.98% on the six-class and four-class tasks, respectively. These findings validate the effectiveness and superiority of M-GGF in modeling complex emotional dynamics, providing a promising direction for next-generation affective computing systems.