Multidimensional Exploration of Influencing Factors of College Students' Academic Performance: An Empirical Study Based on 18 Machine Learning Algorithms

Technical Summary: Multidimensional Exploration of Influencing Factors of College Students' Academic Performance

Problem Statement
Academic performance is a complex outcome shaped by the interplay of individual characteristics, family environment, and school education. While existing literature has extensively explored these dimensions, significant limitations remain. Traditional studies often rely on classical statistical methods (e.g., regression analysis) or single machine learning algorithms, which struggle to capture complex, non-linear correlations among multiple factors. Furthermore, the "black box" nature of many machine learning models hinders interpretability, making it difficult for educational managers to trust model outputs or formulate targeted intervention strategies. There is a critical need for a framework that not only predicts academic performance with high accuracy but also explains the underlying mechanisms of influencing factors to support practical educational interventions.

Methodology
This study employs a comprehensive empirical approach using data from 145 college students (Faculty of Engineering and Faculty of Educational Sciences) from the University of California, Irvine (UCI) dataset. The research integrates 30 indicators covering personal information, family background, and educational status.

1. Data Pre-processing: The target variable, "GRADE," was converted into a binary classification task (0 for "Fail," 1 for "Pass") to address the initial multi-class complexity. To mitigate the severe class imbalance (approximately 17:1 ratio of pass to fail), the Synthetic Minority Oversampling Technique (SMOTE) was applied to generate synthetic samples for the minority class.
2. Model Comparison: The study systematically evaluated 18 machine learning algorithms, including Naive Bayes, Gradient Boosting Machine, XGBoost, Logistic Regression, Neural Networks, Gaussian Process, Bagging Decision Trees, LightGBM, and various linear discriminant and regression models.
3. Evaluation Metrics: Model performance was assessed using Accuracy, AUC (Area Under the Curve), Kappa, Sensitivity, Specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), Precision, Recall, and F1-score. A 5-fold, 10-times cross-validation strategy was utilized to ensure robustness.
4. Interpretability: The LightGBM model, identified as the top performer, was further analyzed using the SHAP (Shapley Additive exPlanations) framework. This method quantifies the contribution of each feature to the model's prediction, allowing for an in-depth exploration of feature importance and directionality.
5. Application Development: An online prediction system was developed using the Shiny framework to visualize model predictions and feature interpretations, facilitating practical use by educators.

Key Results

• Model Performance: Among the 18 algorithms tested, the LightGBM algorithm achieved optimal results. It demonstrated an Accuracy of 0.985, an AUC standard deviation of 0 (indicating high stability), a Kappa of 0.951, Sensitivity of 0.949, Specificity of 1, and an F1-score of 0.974. Other strong performers included Bagging Decision Trees and XGBoost, but LightGBM consistently outperformed traditional models like Classification and Regression Trees (Accuracy 0.896) and Logistic Regression (which showed higher AUC standard deviation).
• Feature Importance: SHAP analysis revealed that personal factors and school education factors are the dominant drivers of academic performance, while family factors (e.g., parental education, work status) have a relatively weak impact.
• Key Significant Variables: The most influential variables identified were:
  • Personal Dimension: Gender, Age, having a partner (HasPartner), and engaging in part-time work.
  • Learning Behavior: Class listening status (ClassListening).
  • Expectations: Expected cumulative average score upon graduation (ExpectedGradCGPA).
• Specific Findings:
  • Students without partners showed better performance predictions than those with partners.
  • Students aged 22–25 demonstrated the best performance predictions, suggesting an optimal learning focus period.
  • High expected graduation grades were associated with negative predictive effects, likely due to anxiety.
  • Part-time work showed a positive predictive effect, potentially linked to improved time management and practical application of knowledge.

Key Contributions

1. Refinement of Influencing Dimensions: The study clarifies the hierarchy of impact, verifying that personal and school education factors dominate over family factors in the college context. It identifies specific, actionable indicators (e.g., class listening, relationship status) rather than broad categories.
2. Methodological Optimization: By comparing 18 algorithms and integrating SHAP interpretation, the study overcomes the "black box" limitation of traditional machine learning. It replaces single-algorithm approaches with a comparative framework that ensures both high prediction accuracy and interpretability.
3. Practice-Oriented Intervention Strategies: Moving beyond macro-level suggestions, the study proposes specific, data-driven interventions based on SHAP insights. These include strategies to alleviate failure anxiety, guide views on romantic relationships, establish lifelong learning concepts, and cultivate practical abilities through industry-education integration.

Significance
The paper claims that its research provides a "data-driven" quantitative decision-making basis for educational managers. By constructing a closed loop from "model selection" to "factor analysis" and "intervention implementation," the study offers a practical paradigm for applying educational data science. The development of the online Shiny application lowers the technical threshold for educators, enabling them to intuitively identify academic risks and tailor personalized, intelligent intervention programs. Ultimately, the study aims to improve college students' academic performance and the overall quality of higher education by bridging the gap between complex predictive modeling and actionable educational policy.