Multi-modal information constraint inference and adaptive algorithm for psychological health risks of university students

Abstract

With the advent of the big data era and the growth of data mining (DM) technology, it has become possible to use DM models to provide real-time perception of the psychological state of university students. The accurate identification of psychological health risks among university students requires the model to fully integrate multimodal information and understand its complex interactions. We have designed a Multi Subspace Sentiment Analysis (MSSA) model to address the shortcomings of existing methods in modality collaboration and semantic depth capture. The model first uses a stacked bidirectional long short-term memory network (BiLSTM) to extract deep features at the discourse level; Furthermore, through a dual projection mechanism, modal invariant representations (learning cross modal commonalities) and modal specific representations (preserving single modal characteristics) are constructed separately; On this basis, a Transformer based self attention fusion module is introduced to explicitly model the collaborative interaction between modalities and capture richer latent semantic information; Finally, the fused joint representation outputs the classification result through the fully connected layer. Simulation experiments show that the MSSA framework outperforms mainstream baseline models in terms of precision and F1 score in identifying mental health risks, and demonstrates good adaptability and robustness.