🎓 Student Performance Prediction and Fairness Analysis

This project explores student academic performance prediction using machine learning, combined with data analytics and bias detection/mitigation techniques.
The goal is to understand how different demographic, academic, and social factors influence student grades — and ensure that the predictive model is fair and unbiased toward protected attributes like gender and past academic failures.

---

🧠 Project Overview

The dataset used here comes from the UCI Machine Learning Repository, containing student data from two Portuguese schools:

• Gabriel Pereira (GP)
• Mousinho da Silveira (MS)

The analysis includes:

• Data preprocessing and ETL
• Exploratory Data Analysis (EDA)
• Feature Engineering
• Logistic Regression model training
• Bias detection using aif360
• Bias mitigation using the Reweighing algorithm

---

🗂️ Dataset Description

File: student-mat.csv
Rows: 395
Columns: 33

Each record represents a student and includes:

• Demographic information: age, sex, address
• Family background: parent jobs, education level
• Academic performance: grades G1, G2, G3
• Social factors: study time, failures, alcohol use, free time
• Final grade (G3) — used to predict student success

---

⚙️ Data Processing and ETL Steps

1. Data Import and Inspection

data = pd.read_csv('student-mat.csv')
data.head()

• No missing values were detected.
• Added a new column Grade as a percentage: [ \text{Grade} = \frac{G3}{20} \times 100 ]

---

2. Data Cleaning

• Verified for null values using a heatmap.
• No imputation required — dataset is complete.
• Renamed or encoded categorical variables for modeling.

---

3. Feature Engineering

• Created a binary variable Result:
  pass if G3 >= 10, otherwise fail.
• Added failed_before = True if failures > 0.
• Performed one-hot encoding for nominal features:
  • Mjob, Fjob, reason, guardian
• Used Label Encoding for binary features:
  • sex, address, famsize, schoolsup, etc.

---

📊 Exploratory Data Analysis (EDA)

Several insights were drawn through visualizations:

🏫 School Distribution

• ~88% of students belong to Gabriel Pereira (GP).
• ~12% belong to Mousinho da Silveira (MS).

👩‍🦰 Gender Ratio

• Female students: 208
• Male students: 187

🏠 Address Type

• Urban: ~72%
• Rural: ~28%

📈 Performance Bias Indicators

• Male students have slightly higher average grades (Mean: 54.57%) than females (Mean: 49.83%).
• Students who have never failed before show much higher performance (Mean: 56.26%) than those who failed before (Mean: 36.32%).

---

🤖 Machine Learning Model

Model: Logistic Regression

A binary classification model was trained to predict whether a student passes or fails (G3 >= 10).

Training-Test Split:

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)

Performance Metrics

Metric	Score
Accuracy	0.916
Precision	0.92
Recall	0.945
F1-score	0.932
MAE	0.08

Confusion Matrix:

[[40  6]
 [ 4 69]]

✅ The model performed well with 91.6% accuracy, showing strong balance between precision and recall.

---

⚖️ Bias Detection & Mitigation

Bias was tested against two protected attributes:

• Gender (sex)
• Previous academic failure

---

1. Bias Detection

Using IBM’s AI Fairness 360 (aif360) toolkit:

from aif360.metrics import BinaryLabelDatasetMetric
metric_orig_train = BinaryLabelDatasetMetric(
    bias_dataset,
    unprivileged_groups=[{'sex': 0}],
    privileged_groups=[{'sex': 1}]
)
metric_orig_train.mean_difference()

📊 Result:

Difference in mean outcomes between unprivileged (female) and privileged (male) = -0.066

→ Indicates slight bias favoring males.

---

2. Bias Mitigation – Reweighing Algorithm

To correct unfair representation:

from aif360.algorithms.preprocessing import Reweighing
RW = Reweighing(unprivileged_groups, privileged_groups)
RW.fit(bias_dataset)
train_tf_dataset = RW.transform(bias_dataset)

📈 Post-Mitigation:

Difference in mean outcomes = 0.000000

✅ Bias successfully removed — fairness achieved between groups.

---

🧾 Key Insights

• Male students performed slightly better than female students in raw data.
• Students with a history of failure had significantly lower grades.
• The trained logistic regression model achieved high accuracy and generalization.
• The Reweighing algorithm from aif360 effectively neutralized observed gender bias.

---

🧰 Tech Stack

Category	Tools / Libraries
Language	Python
Data Handling	Pandas, NumPy
Visualization	Matplotlib, Seaborn
Modeling	scikit-learn
Fairness & Ethics	aif360
Metrics	Accuracy, Precision, Recall, F1, MAE

---

📦 Installation

# Clone this repository
git clone https://github.com/yourusername/student-performance-fairness.git
cd student-performance-fairness

# Create virtual environment
python -m venv .venv
source .venv/bin/activate   # (Linux/macOS)
# or
.venv\Scripts\activate      # (Windows)

# Install dependencies
pip install -r requirements.txt

---

📚 Dependencies

Include these in your requirements.txt:

numpy
pandas
matplotlib
seaborn
scikit-learn
aif360

---

📈 Future Work

• Explore other ML models (RandomForest, XGBoost) for better interpretability.
• Implement Fairness-aware post-processing methods (e.g., Reject Option Classification).
• Extend bias analysis to other features like address, parent education, or internet access.
• Deploy model as a small web dashboard using Streamlit or FastAPI.

---

🧑‍💻 Author

Anas Mustafa
AI/ML Developer | NLP & RAG Systems | Data Science Researcher
📍 Cottbus, Germany
🔗 LinkedIn | GitHub