Application of deep learning algorithms in improving student learning quality assessment of applied psychology micro-major programs

Abstract Applied psychology micro-major programs, as emerging interdisciplinary education models, face challenges in student learning quality assessment including complex evaluation dimensions and heterogeneous learning behavior data. Addressing the problem that existing assessment models struggle to capture deep temporal dependencies and key influencing factors in students’ learning processes, this study constructs a deep learning assessment framework integrating Bidirectional Long Short-Term Memory networks and multi-head attention mechanisms. It is important to clarify that the present work focuses on assessing the quality of student learning outcomes and processes within the program—rather than evaluating instructors’ teaching behaviors or institutional delivery quality—a distinction that is elaborated in the Introduction. The framework introduces a domain-adapted heterogeneous feature embedding strategy that projects multi-source educational data—structured administrative records, behavioral logs, psychometric scales, and text-derived features—into a unified latent space, and employs a residual-connected feature fusion module that deeply couples attention-weighted representations with bidirectional temporal encodings. The framework integrates multi-source data including learning behavior sequences and psychological assessment data, utilizes BiLSTM networks to extract bidirectional temporal features, and combines multi-head attention mechanisms to achieve adaptive weight allocation and feature fusion of assessment elements, establishing a high-precision student learning quality prediction model. Experimental verification based on data from 180 students across two consecutive semesters employs student-level 5-fold cross-validation—ensuring all 15-week observation sequences of the same student remain in the same fold and that all preprocessing steps are fitted exclusively on each fold’s training partition—showing that the proposed model achieves an assessment accuracy of 97.23 ± 0.93%, an improvement of 14.83% points over the weakest traditional baseline (SVM) and 3.08% points over single BiLSTM models, with an inference time of 1.8 milliseconds. Nine baseline models—including SVM, Random Forest, XGBoost, BP Neural Network, single-layer LSTM, single-layer BiLSTM, Transformer Encoder, CNN-BiLSTM, and TabNet—are compared to confirm the architectural advantage of the proposed model over both conventional machine learning, educational AI-specific, and attention-driven alternatives. A systematic ablation study quantifies the individual contributions of the BiLSTM encoder, multi-head attention module, residual feature fusion strategy, and each heterogeneous data source. Statistical significance of all improvements is confirmed via McNemar’s test ( p < 0.05). Model interpretability is established through dual evidence: multi-head attention weight visualization and DeepSHAP-based feature importance analysis, with cross-validated consistency between the two methods. The model can effectively identify core factors affecting student learning quality and provide interpretable assessment evidence, offering precise data support for applied psychology micro-major program reform.