Research on feature extraction and classification algorithm of large-scale psychological evaluation data in universities

With the increasingly prominent mental health problems of college students, the traditional psychological assessment methods are difficult to meet the needs of large-scale data processing because of their strong subjectivity and low efficiency. This paper puts forward an integrated analysis framework of "feature-model-application", aiming at improving the effectiveness of mental health services in universities through multimodal data fusion technology. In feature extraction, a Sparse Autoencoder (SAE) is used for dimensionality reduction and feature enhancement of scale data, for text data, a hierarchical BERT-TextCNN fusion model is constructed to extract semantic features, and for behavioral time-series data, a Sliding Window Dynamic Graph Convolutional Network (SW-DGCN) is designed to capture the evolution of behavioral patterns. In the design of the classification model, a Gated Cross-Attention (GCA) mechanism is employed to integrate multi-modal features, and strategies such as dynamic depthwise separable convolution, knowledge distillation, and adaptive pruning are applied to achieve model lightweighting. In addition, a dynamic early warning system is developed, and online incremental learning mechanism is introduced to adapt to the continuous changes of students' psychological state, and accurate early warning is provided through the dynamic division of risk levels. The experimental results show that the multimodal fusion model proposed in this paper is superior to the baseline model in classification performance, real-time performance and resource consumption. The AUPRC reaches 0.813, the F1 score is 0.782, the parameters are only 48.3MB, and the reasoning time is 42 ms. This study provides an efficient and practical solution for the early warning of psychological crisis in universities, which has important theoretical and application value.