EmoMamFusion: A sentiment classification algorithm framework based on a differentiated dual-branch fusion architecture

Multimodal sentiment recognition enables precise analysis and classification of human emotions. However, existing methods face three major challenges: low feature fusion efficiency due to modality heterogeneity, insufficient long-range dependency modeling capability, and high model complexity. To address these issues, this paper proposes the EmoMamFusion model for multimodal sentiment recognition, based on a differentiated dual-branch fusion architecture. Specifically, the model achieves differentiated fusion through the TranFusion and MamFusion branches: TranFusion, centered around cross-attention, performs preliminary cross-modal semantic alignment and shallow feature fusion; MamFusion models long-range dependencies through the Mamba network and enforces cross-modal collaboration with InfoNCE loss, enabling deep feature fusion and semantic enhancement. Finally, the features from both branches are globally integrated by a Transformer encoder to complete sentiment category prediction. Experiments on the CMU-MOSI and CMU-MOSEI datasets show that EmoMamFusion outperforms 22 state-of-the-art methods in core metrics such as binary classification accuracy (Acc-2), F1 score, seven-class accuracy (Acc-7), and Pearson correlation coefficient (Corr). Specifically, on the CMU-MOSI dataset, Acc-7 and Corr reached 50.96% and 0.895, respectively, while on the CMU-MOSEI dataset, Acc-7 and Corr reached 54.98% and 0.821. At the same time, the model has only 1.38M parameters, achieving a collaborative optimization of high performance and lightweight design.