Optimization and application of efficient multimodal convolutional neural network model for big data analysis

Traditional Convolutional Neural Network (CNN) performs well in processing single modal data, but it still faces many challenges in multi-modal data fusion, such as high model complexity, low training efficiency and insufficient feature extraction. In order to solve the above problems, this paper designs a CNN architecture integrating multimodal features, which processes image and text data respectively through the multimodal feature extraction layer, and realizes the deep interaction and fusion of features in the multimodal fusion layer by using the cross-attention mechanism. In addition, an adaptive hybrid optimizer is proposed, which combines the advantages of Adam and SGD, and adopts gradient clipping strategy and cosine annealing learning rate adjustment method to improve the training efficiency and stability of the model. On the training strategy, the multi-stage training method is adopted. The image branches are pre-trained, and then the whole network is fine-tuned. Dropout and L2 regularization strategies are configured to prevent the model from over-fitting. In this paper, the optimized multi-modal CNN model is applied to the forecasting task of stock price fluctuation trend in financial investment decision-making. By combining the news text of listed companies with the daily K-line chart data, a multi-modal prediction model is constructed. The experimental results show that the model proposed in this paper is superior to the traditional single-modal model and existing multi-modal methods in weighted F1-score and AUC-ROC, and it also shows good performance in reasoning efficiency. In addition, the attention distribution analysis of the model reveals its focus on key information in the decision-making process, which further verifies its effectiveness in multimodal data fusion.