As the diversity and complexity of emotional expressions increase, single-modal emotion recognition is difficult to meet the requirements, and multimodal fusion prediction faces the challenges of computational complexity and modal information integration. This study proposes a machine learning-based multimodal emotion recognition method that fuses text, audio, and video modalities to further improve the prediction accuracy. The text modality uses FastText to extract word vector features, the audio modality uses MFCC to extract acoustic features and the video modality extracts features through ResNet-18 combining spatial and temporal attention mechanisms. This study proposes an innovative multilevel fusion strategy on this basis: feature fusion for text and audio, followed by decision fusion with video modality, and Stacking ensemble learning is used to optimize the decision-making process in the decision fusion part. This method has the best fusion effect comparing with traditional decision fusion methods such as dynamic weighting and entropy weighting. Experiments on the CMU-MOSI dataset show that the method has an accuracy of 87.04% and an F1 score of 0.9014, which significantly outperforms the unimodal and traditional multimodal methods, and demonstrates a strong potential for application in complex scenarios.