🎓 Early Student Failure Detection System

🚀 Machine Learning Semester Project | Dec 2025

Role: ML Engineer & Data Analyst
Tech Stack: Python, Scikit-Learn, Pandas, Seaborn

---

📖 Project Overview

High failure rates in higher education often stem from a lack of timely intervention. Traditional monitoring is reactive—educators only notice struggling students after they fail.

This project implements an Early Warning System using the Decision Tree Classifier. Unlike standard academic projects that assume "perfect data," this implementation focuses on the Real-World Research Gap: handling Data Quality Anomalies (e.g., negative study hours, impossible grades) to build a robust, interpretable model.

---

🚩 The Research Gap & Solution

Most existing literature (e.g., Ahmed et al., 2024) focuses purely on model architecture. However, real-world educational registries are noisy.

🛑 The Problem (Raw Data)	✅ Our Solution (The Pipeline)
Impossible Values: Negative study hours (-5h), Grades > 100%.	Logic-Based Cleaning: Automated anomaly detection & domain-logic correction.
Missing Targets: Students with no Pass/Fail label.	Data Integrity Check: Removal of unlabelled rows to prevent ground-truth corruption.
Class Imbalance: Failing students are often a minority.	Stratified Splitting: Ensuring equal representation of "At-Risk" students in training/testing.

---

🛠️ Implementation Details

1. Data Processing Pipeline

• Anomaly Detection Audit: Automated scan for invalid range values.
• Imputation: Median strategies for numerical gaps, Mode for categorical gaps.
• Feature Engineering: One-Hot Encoding for 'Parent Education' & 'Activities'.

2. Model Architecture

• Algorithm: Decision Tree Classifier (Optimized for Interpretability).
• Hyperparameter Tuning: GridSearchCV used to optimize max_depth (preventing overfitting) and min_samples_split.
• Metric Focus: Prioritized Recall (to minimize False Negatives - i.e., missing a failing student).

---

📊 Key Results & Findings

🏆 Model Performance (On Validated Data)

After switching to the high-quality Student Performance Factors dataset to validate the pipeline:

• Accuracy: ~85% - 92%
• Recall (Fail Class): High sensitivity to detecting at-risk students.

📉 Critical Data Analysis (The "Noise" Discovery)

During the initial phase, we audited a legacy dataset (student_performance_prediction.csv).

• Discovery: A 0.001 correlation was found between Study Hours and Passing.
• Conclusion: We mathematically proved the legacy dataset contained synthetic noise (randomized labels), demonstrating the importance of Data Quality Auditing before modeling.
• (See "Control Experiment" section in the notebook for the visual proof).

---

Repository Structure

├── 📓 Project_Early_Student_Failure_Detection_System.ipynb  # Main Analysis Notebook
├── 📄 StudentPerformanceFactors.csv                         # The High-Quality Dataset
├── 📄 student_performance_prediction.csv                    # (Archived) The Noisy Legacy Data
└── 📝 README.md

---

🚀How to Run

1. Open the .ipynb file in Google Colab or Jupyter Notebook
2. Ensure the .csv dataset is in the same directory
3. Run the cells sequentially to see Data Audit, Training, and Evaluation steps

---

👨‍💻Contributors

• Kaleb Kebede (ID: 1501009) – Lead ML Engineer
• Abel Adam (ID: 1500592) – Data Analyst
• Daniel Abebe (ID: 1501647) – Researcher

---

Submitted to Mr. Mesay A. | Department of Software Engineering