In view of the complexity of emotional expression in multimodal content (text, images, audio, etc.) generated by users on social media, and how these emotional expressions affect consumer behavior, this paper proposes a framework for emotional transformation and consumer behavior prediction based on multimodal learning. This paper first preprocesses text, images, and audio, and extracts features using technologies such as BERT (Bidirectional Encoder Representations from Transformers). Then, further encoding is performed to generate high-level feature representations. Then, sentiment classification and regression are performed, and consumer behavior is predicted through late fusion strategies and multimodal RNN (Recurrent Neural Network), combined with user historical behavior data, and model optimization technology is applied to improve prediction accuracy. In the sentiment classification task, the accuracy and F1 score of the multimodal learning framework are higher than those of the single-modal method. In the consumer behavior prediction task, the average accuracy of the multimodal learning framework is about 95%, which is about 11% higher than the traditional logistic regression, and the prediction time is also reduced by about half. These results fully verify the effectiveness and efficiency of the multimodal learning framework in social media sentiment conversion and consumer behavior prediction.