A reproducible machine learning project that predicts whether a student will Pass or Fail using the UCI Student Performance dataset and compares multiple classification models under two scenarios: with and without previous grades.
This project applies supervised machine learning to predict whether a student will pass or fail a course.
The target variable is engineered from the final grade:
pass_final = 1 if G3 >= 10
pass_final = 0 if G3 < 10
The project evaluates four classification algorithms:
- Logistic Regression
- Decision Tree
- Random Forest
- Support Vector Machine (SVM)
Two experiments are compared:
- With previous grades (
G1,G2) - Without previous grades (
G1,G2) to simulate an earlier-warning setting before prior grade information is available
Source: UCI Student Performance Dataset
Course file: student-por.csv
| Item | Value |
|---|---|
| Student records | 649 |
| Original features | 33 |
| Final project columns | 34 |
| Passed | 549 |
| Failed | 100 |
| Pass rate | ~84.6% |
The features cover demographic, family, school, behavioral, support-related, and academic information.
UCI Student Performance Data
│
▼
Data Inspection
│
▼
Cleaning & Target Creation
│
▼
Exploratory Data Analysis
│
▼
Preprocessing Pipeline
├─ Numerical: Imputation + Scaling
└─ Categorical: Imputation + One-Hot Encoding
│
▼
Train/Test Split
80% / 20%
│
▼
Four Classification Models
│
▼
Accuracy • Precision • Recall
F1 • ROC-AUC
│
▼
Model Comparison & Interpretation
Uses G1 and G2 as predictors.
This scenario answers:
How accurately can final pass/fail performance be predicted when previous period grades are already available?
Excludes G1 and G2.
This scenario is more relevant to an early-warning system, because the model must rely on demographic, behavioral, support, attendance, and other non-grade factors.
| Model | Accuracy | Precision | Recall | F1 | ROC-AUC |
|---|---|---|---|---|---|
| Logistic Regression | 0.9077 | 0.9455 | 0.9455 | 0.9455 | 0.9273 |
| Random Forest | 0.9000 | 0.9292 | 0.9545 | 0.9417 | 0.9309 |
| SVM | 0.8769 | 0.8983 | 0.9636 | 0.9298 | 0.9168 |
| Decision Tree | 0.8538 | 0.9174 | 0.9091 | 0.9132 | 0.7295 |
| Model | Accuracy | Precision | Recall | F1 | ROC-AUC |
|---|---|---|---|---|---|
| SVM | 0.8308 | 0.8607 | 0.9545 | 0.9052 | 0.6691 |
| Random Forest | 0.8000 | 0.8500 | 0.9273 | 0.8870 | 0.6557 |
| Logistic Regression | 0.7923 | 0.8673 | 0.8909 | 0.8789 | 0.6227 |
| Decision Tree | 0.7846 | 0.8596 | 0.8909 | 0.8750 | 0.5455 |
Logistic Regression with G1 and G2
- Accuracy: 0.9077
- Precision: 0.9455
- Recall: 0.9455
- F1-score: 0.9455
- ROC-AUC: 0.9273
SVM without G1 and G2
- Accuracy: 0.8308
- F1-score: 0.9052
- Recall: 0.9545
This shows that useful early-warning signals remain available even when previous grades are excluded.
G1andG2are the strongest predictors of final performance.- Previous failures are an important negative signal.
- Without prior grades, variables such as failures, absences, higher-education intention, family relationship quality, weekend alcohol consumption, going out, age, and school become more influential.
- Because the dataset is imbalanced toward passing students, F1-score and recall are more informative than accuracy alone.
- Median imputation for numerical features
- Standard scaling using
StandardScaler - Most-frequent imputation for categorical features
- One-hot encoding with unknown-category handling
ColumnTransformer+Pipelinefor reproducibility
Test size: 20%
Random state: 42
Stratified split: Yes
Logistic Regression
Decision Tree
Random Forest
SVM (RBF kernel)
Student-Performance-Prediction/
├── README.md
├── student_performance_prediction.ipynb
├── student_performance_prediction.py
├── requirements.txt
├── cleaned_student_performance.csv
├── model_results.csv
├── model-comparison-preview.png
├── Student_Performance_Data_Mining_Report.pdf
└── Student_Performance_Data_Mining_Presentation.pdf
- The Pass class is much larger than the Fail class.
- The dataset represents a specific educational context and may not generalize directly to other schools or countries.
- Using
G1andG2improves predictive performance but reduces the value for very early intervention. - Behavioral and social attributes require careful consideration of privacy, fairness, and ethical use.
- Hyperparameter tuning and cross-validation could further strengthen the evaluation.
- Cross-validation
- Hyperparameter tuning
- Fairness analysis
- Class-imbalance techniques
- Explainability / feature interpretation
- Deployment as an academic early-warning dashboard
Mustafa Abughareebeh
Data Science & Artificial Intelligence Undergraduate