Deep learning-based classification of student GPA integrating psychological and family factors in the post-pandemic era

Key findings from this study

• The study found that the TabTransformer with feature-gating achieved the highest classification accuracy (0.798) among evaluated deep learning models for GPA prediction.
• The authors report that GPA was significantly negatively correlated with depression and anxiety as measured by SCL-90 assessment.
• The researchers demonstrate that unfavorable family factors such as economic disadvantage and prolonged parental separation correlated with poorer psychological outcomes.

Overview

This study developed a deep learning framework integrating family background and psychological factors to classify undergraduate GPA in the post-pandemic context. The research addresses limitations of conventional prediction models by incorporating non-cognitive dimensions alongside academic records. Data encompassed 1,692 Chinese undergraduates with assessments of family circumstances, psychological evaluation scores, and academic performance.

Methods and approach

The researchers collected data including family background variables (gender, economic status, only-child status, years of parental separation) and SCL-90 psychological evaluation scores alongside GPA records. Four deep learning architectures were compared: TabTransformer, DCNv2, AutoInt, and MLP-ResNet. A lightweight feature-gating mechanism enhanced feature selection within high-dimensional heterogeneous data. Model evaluation utilized Accuracy and Area Under the ROC Curve metrics. Statistical associations were examined through Spearman's rank correlation, chi-squared tests, and t-SNE visualization.

Results

The TabTransformer model with gating mechanism demonstrated superior performance, achieving Accuracy of 0.798 and AUC of 0.833 compared to alternative architectures. GPA exhibited significant negative correlations with multiple SCL-90 domains, particularly depression and anxiety measures. Unfavorable family circumstances—including lower economic status and extended periods of parental absence—associated with diminished psychological assessment outcomes and reduced academic performance.

Implications

The framework enables systematic identification of students at academic risk through quantifiable family and psychological indicators. Integration of non-cognitive factors into predictive models enhances classification accuracy for targeted intervention design. Results support institutional implementation of early-warning systems that incorporate psychological dimensions beyond traditional academic metrics.

Findings indicate that psychological distress mediates relationships between family adversity and academic outcomes. Institutions implementing this classification approach could stratify students for differentiated support services. The methodology demonstrates feasibility of deploying deep learning models in educational contexts with heterogeneous data types.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.