Student Burnout Risk Prediction: Practical Analysis of a Production-Grade Machine Learning Pipeline

This project is an end-to-end production-ready machine learning project focusing on predicting student burnout risk levels. Core technologies include LightGBM regressor and custom threshold optimization strategy, covering the full workflow of feature engineering, overfitting mitigation, and FastAPI microservice deployment. The project is sourced from GitHub user R-Harieharan's student-burnout-api, with data from the Kaggle Student Performance and Burnout Dataset (50,000 records).

Student burnout is a focus of attention in contemporary education, especially against the backdrop of the popularization of generative AI tools. This project aims to identify burnout risk levels based on student behavior data and learning patterns. Unlike proof-of-concept or competition models, it is a complete production-grade solution covering the entire engineering practice from data preprocessing to model deployment.

The project uses LightGBM as the core algorithm. Facing the bottleneck of approximately 48% validation set accuracy with the standard multi-classification architecture, it breaks the limitation by converting the classification problem into a regression problem via ordinal mapping. To address the overfitting phenomenon (100% accuracy on training data), the LightGBM sequential gradient boosting architecture is introduced, reducing the training accuracy to 55% to maintain generalization ability.

The project designs a layered preprocessing framework: ordinal variables (e.g., academic year, skill level) use ordinal mapping to preserve order; other features implement custom Scikit-Learn estimators, including dynamic interaction feature calculation (e.g., GPA change over time) and outlier clipping layers; through Recursive Feature Elimination (RFE), features are compressed to 7 high-impact features, improving interpretability and inference speed.

After converting classification to regression, stable decision thresholds are determined via grid search: low→medium is 0.643, medium→high is 1.271, which improves the recall rate of high-risk groups (meeting educational intervention needs). The final model achieves an overall test set accuracy of 50.74%, medium-risk recall rate of 67%, high-risk precision rate of 65%, and macro-average F1 score of 0.50, showing robust performance in high-noise three-classification tasks.

The project implements a complete FastAPI backend that can load preprocessing pipeline states, handle out-of-vocabulary data anomalies, and perform sub-second real-time inference. The containerized architecture facilitates integration into student management systems. Engineering insights include: problem refactoring (classification to regression), feature engineering prioritized over complex models, production systems needing to focus on reusable pipelines, anomaly handling, and clear API design.

The project provides an early warning system for educators, helping to allocate counseling resources rationally and improve student well-being. In the future, it is necessary to balance technical capabilities and privacy protection to ensure the tool serves student growth rather than monitoring. With the improvement of educational data and the popularization of ML, similar application scenarios will increase.