SHAP-Weighted Expert Fusion for Multimodal Emotion Recognition

The study evaluates XGAF, a mixture-of-experts approach using TreeSHAP attribution to weight multimodal inputs for sentiment recognition. Results show that using sum-abs reduction prevents high-dimensional experts from being suppressed, allowing the Transformer variant to reach 0.5983 F1 on the MELD dataset.