This chapter presents a comprehensive and explainable multimodal sentiment analysis framework that systematically compares early and late fusion strategies using three language encoders (BERT, DistilBERT, and RoBERTa) and three vision encoders (ResNet50, VGG16, and ViT). We investigate 12 model configurations by combining one another under two classification backbones: XGBoost and LightGBM. Experimental results on the MVSA-Single dataset reveal that early fusion models consistently outperform their late fusion counterparts in both accuracy and F1-score. Notably, configurations leveraging ViT and BERT or RoBERTa exhibit superior performance and interpretability. Explainable AI (XAI) tools such as SHAP are employed to visualize modality contributions, revealing that early fusion facilitates better modality alignment and feature integration. The findings underscore the potential of early fusion for explainable and effective sentiment modeling in multimodal settings.