FYP Research Poster Summary
Based on the original Final Year Project (FYP) poster titled "Predicting Student Academic Performance" by Muhammad Ilham Bin Noor Effendy (Student ID: 2022674776 | Portfolio):
- Problem Statement: Academic failures are often only detected reactively during final grading. Early identifier models are critical to supply educators with predictive alerts for proactive academic mentoring.
- Project Objective: To build, benchmark, and deploy predictive models using student demographics, home environment, and study profiles to forecast academic grades.
- Baseline Models: Benchmarked three classifiers: Support Vector Machine (SVM), Backpropagation Neural Network (BPNN), and K-Nearest Neighbors (KNN) to evaluate prediction boundaries.
Dataset Preprocessing & Adaptation
To upgrade the model from the original study, the data processing pipeline was thoroughly refined:
- Multi-Source Data Merging: Combined math and language profiles (674 students) into a consolidated repository, cleaning overlapping profiles by matching attributes (age, gender, address, family status, etc.).
- Cultural Localization: Dropped irrelevant variables (e.g., daily and weekly alcohol consumption: `Dalc` & `Walc`) to align variables with Malaysian secondary school contexts.
- SPM Grade Translation: Mapped final outcomes to the Malaysian Sijil Pelajaran Malaysia (SPM) letter grading system:
- Score 18-20: A+ | Score 16-17: A | Score 15: A-
- Score 14: B+ | Score 13: B | Score 12: C+ | Score 10-11: C (Credit Target)
- Score 9: D | Score 8: E | Score 0-7: G (Fail)
- Feature Pipeline: Applied standard scaling ($\mu$-subtracted, unit-variance scaled) to numeric inputs and drop-first One-Hot Encoding to categorical variables to remove multicollinearity.
Model Optimization & Tuning
The deployed model features significant improvements over basic classifiers:
- Ensemble Architecture: Upgraded to an optimized Random Forest (RF) classifier to reduce model variance.
- Hyperparameter Optimization: Performed a 5-fold cross-validated grid search (2,700 fits) tuning tree estimators, max depth, minimum split samples, and criteria.
- Imbalance Correction: Utilized weighted class penalization to handle skewed pass ratios, elevating the critical F1-Macro score (detecting students at risk of failing to reach a credit grade) to 0.6049.