Student Performance Prediction using Machine Learning

Every term, a predictable fraction of students fails — and most of them were already showing the signs weeks earlier: rising absences, a weak first-period grade, past failures accumulating. Teachers sense this intuitively but cannot quantify it across hundreds of students. This project turns that intuition into a model: trained on the UCI Student Performance dataset (649 students, 33 attributes), gradient-boosted trees predict the final grade G3 from everything known after the first period. The honest headline finding — which the project reports rather than hides — is that past grades dominate the prediction; study time, failures and absences add the rest. The deliverable is positioned carefully as decision support for educators: it flags students for mentoring, never auto-grades or auto-penalizes anyone.

1. The UCI student-mat and student-por datasets are merged (649 records, 33 attributes), cleaned and encoded; G3 is the regression target.
2. Exploratory analysis computes correlations, distributions and the at-risk tail that motivates the early-warning use case.
3. Gradient-boosted trees (with linear-model and random-forest baselines) are trained with cross-validation; R² and MAE are reported.
4. Feature importances and partial-dependence views show what the model learned, checked against the published literature.
5. The predictor demo wraps an interpretable model aligned with the published correlations for instant, explainable estimates.
6. A binned pass/fail classifier is evaluated for the early-warning operating point (predicted G3 < 10 → mentoring flag).

Tech stack:

• Python 3, scikit-learn, XGBoost/LightGBM
• pandas, NumPy
• Matplotlib, Seaborn-style plots
• HTML/CSS/JavaScript predictor demo
• UCI Student Performance dataset

Dataset & model details

• Dataset: UCI Student Performance (Cortez & Silva, 2008) — 649 secondary-school students (395 mathematics, 254 Portuguese), 33 attributes covering demographics, family background, study habits and school support; target G3 on a 0–20 scale.
• Task: Regression of final grade G3 (plus binned pass/fail classification for the warning use case); input = 33 attributes, output = predicted grade.
• Model: Gradient-boosted trees with linear and random-forest baselines; cross-validated.
• Metrics: R² 0.82, MAE ~1.4 grade points (design targets for the built-to-order training run); reported per subject. No measured figures are claimed before the run.

• [Final-grade predictor] Enter study time, absences, past grades and background factors; get a predicted G3 (0–20) with a risk band and plain-language guidance.
• [Per-prediction explanations] The demo shows which factors pushed each prediction up or down, so a counsellor sees reasons, not a black-box number.
• [Correlation analysis] Published correlations of every attribute with G3, visualized — G1/G2 dominate, demographics trail, reported honestly.
• [Early-warning framing] Flags students predicted below the pass mark after the first period, with a documented intervention workflow for schools.
• [Grade-distribution study] The long left tail of failing students is analyzed as the target population for the warning system.
• [Regression + pass/fail modes] Both the G3 regression and a binned at-risk classifier are trained and compared in the notebook.
• [Ethics section] The report explicitly scopes the model as mentoring support and documents misuse boundaries.

• Data cleaning & exploratory-analysis notebook
• Model training & comparison notebook (GBM vs baselines)
• Interactive grade-predictor web demo
• Correlation, distribution and importance plots
• Project report PDF (background, dataset, methodology, results, ethics)
• PPT presentation for final review
• Viva Q&A preparation document (regression metrics, boosting, leakage, ethics of educational ML)

• R² and MAE are design targets; the report records the actual metrics from the training run for the order.
• 649 students from two Portuguese schools — generalization to other systems needs local retraining, stated openly.
• Past grades dominate the prediction; the model cannot foresee sudden personal crises that change a student's trajectory.
• The demo's interpretable model is illustrative; the shipped notebook trains the full gradient-boosted model on the real data.
• Positioned strictly as mentoring decision support — the report forbids using it for grading, streaming or disciplinary decisions.