Student Performance Predictor

An end-to-end machine learning pipeline for predicting student academic performance using structured educational data.

This project explores both classical machine learning and deep learning approaches on the UCI Student Performance dataset, with a strong focus on preprocessing, feature engineering, model comparison, and evaluation.

Live Demo Link : Try it here!

---

Project Overview

The objective is to predict whether a student will pass or fail based on demographic, academic, and behavioral attributes.

Two modeling scenarios were explored:

1. With prior grades (G1, G2)

   • Easier prediction setting with strong academic history signals.

2. Without prior grades (G1, G2 removed)

   • Harder and more realistic prediction setting.

This comparison helps evaluate model robustness under different feature availability conditions.

---

Dataset

Dataset: UCI Student Performance Dataset

• Source: Kaggle / UCI Repository
• Samples: 395
• Features: student demographics, family background, study habits, academic history, lifestyle attributes

Target variable:

• Binary classification:
  • 1 = Pass
  • 0 = Fail

---

Pipeline Architecture

1. Data Loading & EDA

Performed exploratory data analysis to understand feature distributions and dataset characteristics.

Tasks completed:

• Dataset inspection using:
  • head()
  • shape
  • info()
  • describe()
• Missing value analysis
• Duplicate check
• Class imbalance analysis
• Histograms
• Boxplots
• Correlation heatmap
• Outlier inspection

Key observations:

• No missing values
• No duplicates
• Moderate class imbalance
• Strong correlation of G1, G2 with final grade

---

2. Data Preprocessing

Preprocessing steps included:

• One-hot encoding for categorical features
• Standardization using StandardScaler
• Train-test split (80/20)
• Stratified sampling for label balance preservation

Scaling applied only where required:

• Logistic Regression
• SVM
• MLP

No scaling used for:

• Random Forest

---

3. Feature Engineering

Custom engineered features were created to improve predictive performance.

Features engineered:

• study_efficiency
• risk_score
• study_discipline

Purpose:

• Capture behavioral and academic patterns beyond raw features.

Feature engineering was validated through Random Forest feature importance analysis.

---

Models Implemented

1. Logistic Regression

Used as a linear baseline classifier.

Concepts explored:

• linear decision boundaries
• classification probabilities
• regularization

---

2. Support Vector Machine (SVM)

Kernel-based nonlinear classifier.

Hyperparameter tuning:

• C

Concepts explored:

• margin maximization
• nonlinear decision boundaries
• kernel methods

---

3. Random Forest

Ensemble tree-based classifier.

Hyperparameter tuning:

• n_estimators

Concepts explored:

• bagging
• ensemble learning
• feature importance
• variance reduction

---

4. Multi-Layer Perceptron (MLP)

Deep learning benchmark built using TensorFlow/Keras.

Architecture:

• 2 hidden layers
• ReLU activation
• Dropout regularization
• Adam optimizer

Training techniques:

• Early stopping
• Validation split

Concepts explored:

• backpropagation
• gradient descent
• neural network regularization

---

Evaluation Metrics

Models were evaluated using:

• Accuracy
• Precision
• Recall
• F1 Score
• Confusion Matrix

Additional validation:

• 5-Fold Cross Validation
• Learning Curves
• Bias-Variance Analysis

---

Results

Model	Accuracy (With Grades)	F1 (With Grades)	Accuracy (Without Grades)	F1 (Without Grades)
Logistic Regression	0.8734	0.9020	0.6456	0.7544
SVM	0.8228	0.8679	0.6835	0.7934
Random Forest	0.8861	0.9109	0.6709	0.7797
MLP	0.8481	0.8846	0.6709	0.7969

---

Cross Validation Results

Model	CV F1 (With Grades)	CV F1 (Without Grades)
Logistic Regression	0.9333	0.7818
SVM	0.8979	0.7974
Random Forest	0.9354	0.7890

---

Key Findings

With prior grades

Best model: Random Forest

Why:

• highest F1 score
• highest CV F1
• lowest variance across folds

Insight:

• prior academic grades are highly predictive.

---

Without prior grades

Most reliable model: SVM

Why:

• best CV F1 under reduced feature setting

Insight:

• nonlinear models perform better when strong grade predictors are removed.

---

Deep learning insight

MLP did not outperform Random Forest on this small tabular dataset.

Conclusion:

• neural networks are not automatically superior for structured data.

---

Learning Curve Analysis

Learning curves were generated for:

• Logistic Regression
• Random Forest

Insights:

• Logistic Regression showed low bias and mild variance
• Random Forest showed near-perfect training performance with controlled overfitting

MLP was analyzed using:

• training loss curves
• validation loss curves
• early stopping behavior

---

Limitations

• Small dataset size (395 samples)
• Moderate class imbalance
• Strong grade-based predictors simplify task
• Limited hyperparameter tuning

---

Future Improvements

Possible extensions:

• GridSearchCV / RandomizedSearchCV
• SMOTE / class balancing
• Feature selection methods
• XGBoost / LightGBM benchmarking

---

Tech Stack

• Python
• Pandas
• NumPy
• Matplotlib
• Seaborn
• Scikit-learn
• TensorFlow / Keras
• Google Colab

---

Repository Structure

student-performance-predictor/
│
├── data/
├── notebooks/
├── results/
├── README.md
└── requirements.txt

---

Application Demo

---

Run Locally

pip install -r requirements.txt
streamlit run app.py

---

Author

Built as a hands-on machine learning project to strengthen understanding of:

• supervised learning
• preprocessing pipelines
• model training
• evaluation metrics
• cross-validation
• deep learning fundamentals
• bias-variance tradeoff