Student Performance Predictor

A machine learning project that predicts whether a student is likely to pass or fail based on academic and lifestyle factors such as attendance, study hours, previous exam scores, sleep, parental support, and extracurricular activities.

The project includes a complete machine learning pipeline covering data preprocessing, exploratory data analysis (EDA), model training, evaluation, and visualization.

Pipeline

1. Load data — uses a realistic synthetic dataset (1,000 students) so the whole pipeline runs end-to-end with no external file needed. Swap in a real CSV (e.g. the UCI Student Performance dataset) by passing csv_path to load_data().
2. Preprocess — fills missing values, label-encodes categorical features (parental support, internet access, extracurriculars).
3. EDA — saves a feature correlation heatmap and a score distribution plot.
4. Train models — Logistic Regression and Random Forest classifiers, compared on accuracy, ROC-AUC, and 5-fold cross-validation.
5. Evaluate — confusion matrix and feature importance plots for the winning model.
6. Save — best model and scaler are pickled with joblib for reuse.
7. Predict — example inference on a new, unseen student profile.

Results (this run)

Model	Accuracy	ROC-AUC	CV Mean
Logistic Regression	0.840	0.873	0.819
Random Forest	0.830	0.850	0.818

Best model: Logistic Regression. previous_score and study_hours are the strongest predictors of final_score and passed (see correlation_heatmap.png and feature_importance.png).

(Since the dataset is randomly generated each run, exact numbers will vary slightly run to run — set a real CSV via csv_path for reproducible, meaningful results.)

Setup

pip install pandas numpy scikit-learn matplotlib seaborn joblib
python student_prediction.py

Output files

Running the script produces:

• correlation_heatmap.png — feature correlations
• score_distribution.png — distribution of final scores
• confusion_matrix.png — confusion matrix for the best model
• feature_importance.png — Random Forest feature importances
• best_model.pkl — trained model, saved with joblib
• scaler.pkl — fitted StandardScaler, saved with joblib

Using a real dataset

Replace the synthetic generator with your own CSV:

df_raw = load_data(csv_path="your_dataset.csv")

Your CSV should include a final_score (numeric) and/or passed (0/1) target column, plus whatever feature columns you have — the pipeline auto-detects numeric vs. categorical columns during preprocessing.

Reusing the saved model

import joblib
model = joblib.load("best_model.pkl")
scaler = joblib.load("scaler.pkl")  # only needed if the best model was Logistic Regression

Tech stack

Python · pandas · NumPy · scikit-learn (Logistic Regression, Random Forest) · matplotlib · seaborn · joblib