Research Example: Predicting Student Success

- A study by M. Yağcı (2022) examined how machine learning can predict academic performance.
- Researchers collected data from 1,854 undergraduates taking the Turkish Language-I course at a Turkish university.
- Variables included:
- Mid-term exam grades
- Department
- Faculty
- The goal was to predict who would pass the final exam (score of 60 or above).
- Machine learning models used: Random Forest, SVM, k-Nearest Neighbour, Logistic Regression, Naïve Bayes.
- Results showed 70–75% accuracy in predicting final exam success based on these inputs.
- Main finding: Mid-term grades are a significant indicator of final performance, suggesting potential for early identification of at-risk students.

What EDM Can Reveal

- Identifies Key Predictors:
- Even a few data points (mid-term grades, department, faculty) can reliably forecast outcomes.
- Enables Early Intervention:
- Mid-term results give educators a window (about 2.5 months) to support students before finals.
- Uncovers Data Patterns:
- Analysis reveals trends and associations that might otherwise go unnoticed.
- Supports Data-Driven Decisions:
- Insights from EDM can inform interventions, allocate resources, and shape policy at both individual and group levels.
- Promotes Personalized Education:
- EDM contributes to tailoring learning experiences based on real student data, moving beyond generic approaches.

What EDM Cannot Reliably Show

- Correlation vs. Causation:
- Predictive relationships do not mean one factor causes another; other influences like motivation or support may not be visible in the data.
- Missing Context:
- Studies often lack information on demographics, socio-economic status, behavior, or personal circumstances, limiting understanding of student performance.
- Potential for Bias:
- Models can unintentionally reflect or reinforce biases, especially if sensitive factors are included or correlated with other variables.
- Limited Generalizability:
- Findings from a single course or institution may not apply elsewhere; model accuracy can drop with different data or settings.
- Overlooked Human Factors:
- Important elements such as engagement, teacher impact, and emotional well-being are hard to quantify and often omitted from models.
- Risk of Overfitting:
- Machine learning models can identify patterns that do not generalize, especially with complex algorithms or small datasets.

Summary and Considerations

- EDM can effectively highlight students who may be at risk, providing educators with actionable insights based on mid-term grades and departmental context.
- While these models offer valuable early warnings, they do not explain underlying causes or guarantee solutions.
- To fully support student success, EDM findings should be combined with qualitative insights and a broader understanding of the factors influencing learning outcomes.