Empirical study on attention-enhanced deep learning for multimodal big data fusion analysis

To address the limitations in analytical accuracy caused by the strong heterogeneity of multimodal big data and insufficient correlation modeling between modalities, we propose an attention-enhanced deep learning-based multimodal fusion analysis method. This approach constructs a customized multi-level feature extraction network to achieve deep feature mining for heterogeneous data including text, images, and audio. We design a cross-modal adaptive attention fusion mechanism that dynamically learns correlation weights between different modalities, addressing the rigidity of weight allocation in traditional fusion methods. By adopting a multi-objective end-to-end optimization strategy, the method balances classification performance with generalization ability. Experimental validation on three benchmark datasets-MOSI, MOSEI, and MM-IMDb-demonstrates that the proposed method achieves 12.1%higher accuracy than traditional single-modal approaches and 2.2% to 2.7% improvement over mainstream deep learning fusion methods(MulT), with an F1 score of 0.883. This demonstrates significant superiority in both sentiment analysis and movie gen reclassification tasks. The research findings provide a new technical pathway for efficient analysis of multimodal bigdata, offering important theoretical value and promising engineering applications.