Predicting Student Dropout Risk Using Classical Machine Learning: A Complete End-to-End Project

This article introduces a student dropout prediction system based on classical machine learning developed by the Fundació URV's AI fundamentals course. It covers the entire workflow from problem definition and data collection to model deployment, supports the comparison of four algorithms (Logistic Regression, Random Forest, XGBoost, and SVM), and aims to identify dropout risks early to promote educational equity and resource optimization.

Student dropout is a long-standing challenge in the education sector. This project uses binary classification in supervised learning to predict whether a student will drop out. Key decisions are clarified: the learning type is supervised learning (using labeled historical data), the task type is binary classification (dropout/continue studying), and success metrics include accuracy, precision, recall, F1 score, and ROC-AUC. A clear problem definition is the foundation of the project's success, avoiding technical decisions deviating from practical value due to ambiguous definitions.

Seven core features are designed: age (15-25 years old), attendance rate (0-100%), average grade (0-5 points), weekly study duration (0-8 hours), family income (low/middle/high), family support (yes/no), covering demographic, academic performance, and family background. Four classical algorithms are compared:

• Logistic Regression: Strong interpretability, fast training, but difficult to capture non-linear relationships
• Random Forest: Handles complex data, resists overfitting, outputs feature importance, but long training time
• XGBoost: Excellent prediction performance, captures high-order interactions, but prone to overfitting and complex parameter tuning
• SVM: Good performance in high-dimensional space, flexible kernel functions, but weak interpretability and slow training on large-scale data A multi-model comparison strategy helps select the most suitable algorithm.

The Python tech stack is used: Streamlit (for building interactive web interfaces), Scikit-learn (ML algorithms), Pandas (data processing), Matplotlib/Seaborn (visualization). The code is modularly designed: app.py (main application), model.py (model class), data_preprocessing.py (preprocessing), etc. The application includes four modules:

1. Project Overview: Displays background and features
2. Data Exploration: Upload CSV, generate visualizations, statistical information
3. Model Training and Evaluation: Select model, configure dataset ratio, train and visualize metrics
4. Real-time Prediction: Input student features to get risk and intervention suggestions The end-to-end interactive experience allows users to understand the complete ML workflow.

Five metrics are used to evaluate the model:

• Accuracy: Proportion of correct predictions (suitable for balanced classes)
• Precision: Proportion of true dropouts among predicted dropouts (avoids false positives)
• Recall: Proportion of true dropouts correctly identified (more important in education scenarios, avoids false negatives)
• F1 Score: Harmonic mean of precision and recall
• ROC-AUC: Model's ability to distinguish between classes Project value: Early warning (identify high-risk students at the start of the semester), resource optimization (targeted tutoring), policy support (data-driven decision-making), educational equity (help disadvantaged students). It emphasizes that technology should be combined with humanistic care to understand students' difficulties and provide personalized help.

The project is open-sourced under the MIT license, and community contributions are welcome. Expansion directions: Introduce mental health/social relationship features, try deep learning models, develop mobile applications, integrate school information systems. It is an excellent learning resource for ML beginners, showing the complete workflow from data collection to deployment, with a clear code structure suitable for course projects or practice.