Predicting student mathematics performance using machine learning models and identifying key academic predictors.
This project investigates whether machine learning models can accurately predict student Mathematics scores using prior academic performance and school-related metadata, and which subjects influence math outcomes the most.
I worked with a dataset of 16,824 student records from schools in Halabja across five academic years (2018–2023), containing grades in multiple subjects and basic academic metadata.
The project compares Linear Regression and Random Forest Regression, and uses feature importance analysis to understand which subjects are most strongly linked to Mathematics performance.
- Source: Public student grade dataset from schools in Halabja (2018–2023)
- Size: 16,824 records
- Attributes (20 total):
- Metadata: Serial number, School Name, Class, Academic Year
- Subjects: Kurdish, Arabic, English, Mathematics, Physics, Chemistry, Biology, Computer, History, Geography, Economics, Sociology, Philosophy, Science
The target variable is the Mathematics grade.
All other subject grades and metadata are used as input features.
Preprocessing steps:
- Categorical encoding:
- School Name, Class, and Academic Year were encoded using one-hot encoding.
- Missing values:
- Missing numeric values were imputed using the mean of each column.
- Train-test split:
- 80% of the data for training, 20% for testing.
I implemented and compared:
- Linear Regression
- Baseline model assuming a linear relationship between features and math scores.
- Random Forest Regression
- Ensemble of decision trees that can capture nonlinear relationships and interactions between subjects.
Models were evaluated with:
- Mean Absolute Error (MAE)
- Root Mean Squared Error (RMSE)
These metrics measure how far predictions are from actual Mathematics scores (in points).
From the experiments:
| Model | MAE | RMSE |
|---|---|---|
| Linear Regression | 7.49 | 9.54 |
| Random Forest | 7.21 | 9.28 |
- Random Forest achieved lower prediction error (both MAE and RMSE) than Linear Regression.
- This confirms that nonlinear methods like Random Forest can better capture the relationships between school subjects and Mathematics performance.
Random Forest feature importance analysis showed that:
- English
- Science
- Kurdish
were among the most influential subjects for predicting Mathematics scores.
This suggests that:
- language and reading-related skills (e.g., English)
- and scientific reasoning (Science)
may support better performance in Mathematics, for example through improved problem understanding and logical thinking.
This project helped me practice:
- working with a real-world educational dataset
- handling missing values and categorical encoding
- training and comparing regression models
- evaluating models with MAE and RMSE
- using feature importance to interpret Random Forest models
- thinking critically about how data is used in educational decision-making
It also showed me how effort, context, and subject interdependencies all play a role in student performance, not just “raw ability”.
The main experiment notebook is available here:
notebooks/student_math_scores_ml.ipynb
It includes:
- data loading and preprocessing
- training Linear Regression and Random Forest models
- computing MAE and RMSE
- extracting and saving feature importances
The full write-up of this project, including references and figures, is available here:
This document follows a formal structure (Abstract, Introduction, Methodology, Results, Conclusion) and cites related work on student performance prediction using machine learning.
Possible extensions include:
- trying additional models (e.g., Gradient Boosting, XGBoost)
- adding socio-economic or behavioral variables
- testing generalization across other regions or school systems
- exploring fairness and bias to ensure that predictions do not disadvantage specific student groups
Author: Mayssae El Bazi