<p>The growing relevance of emotion recognition has sparked widespread interest, particularly in areas such as psychological assessment, interactive technologies, and robotics. However, current problems in emotional recognition encompass issues like data sparsity, high dimensionality, and the challenges associated with accurately capturing intricate emotional states arising from multi-modal interactions. To improve both accuracy and reliability, it is essential to create a multi-modal fusion strategy that utilizes a variety of data types while effectively addressing these challenges. Therefore, this research introduces a novel model, the Cross Attention Auto-encoded Puma Graph Embedding Model (CAAPGEM), with Puma Optimizer (PO) designed to effectively fuse multi-modal data, thereby enhancing emotional recognition accuracy and addressing existing challenges in capturing complex emotional states. The model employs a Multi-head Self-Attention Transformer Model (MSATM) as the fusion method. This approach adeptly captures complex relationships between visual, audio, and textual data, significantly improving the performance of CAAPGEM in emotional recognition tasks across various contexts. The model's performance is reflected by its F1-score of 99.53%, Acc-2 (binary accuracy) of 99.72%, and Acc-7 (multi-class accuracy) of 99.14%. Also, this model achieves a mean absolute error of 0.525 and a correlation coefficient of 0.932. These results indicate high precision, robust accuracy across different classification levels, and strong correlation, showcasing the model’s effectiveness.</p>