Semisupervised Hierarchical Subspace Learning Model for Multimodal Social Media Sentiment Analysis

The multimodal data analysis model combined with text and image has gradually become an important approach for sentiment analysis in social media. This study proposes a semisupervised hierarchical subspace learning (SHSL) model to address the issue of insufficient labeled samples in multimodal sentiment analysis. The SHSL model captures potential feature representations of multimodal data in a low-rank subspace, at the same time, it adaptively assigns a weight to each modality. As a result, multimodal data can share the potential representation in the low-rank subspace. The SHSL model continuously projects the shared potential representation into the semantic space and achieves label propagation, to link shared potential representations with emotional states in the semantic space. The low-rank subspace serves as a bridge between the original space and the semantic space. It not only enriches the structure of feature space, but also reconstructs original high-dimensional data from low-dimensional features. In addition, the SHSL model constrains the class labels of unlabeled data to satisfy the non-negativity and normalization properties of rows to improve the model performance. Comparative experiments are conducted on the MVSA-single and MVSA-multiple datasets, and the experimental results demonstrate that the proposed model has excellent sentiment analysis capabilities.