Enhancing data pipelines for forecasting student performance: integrating feature selection with cross-validation.

Area of Science:

• Educational data mining
• Predictive modeling in higher education

Background:

• Prior studies focused on data mining methods (DMMs) for student success forecasting.
• Limited research quantified the stability of feature selection techniques in educational contexts.
• Identifying relevant features enhances model interpretability and guides faculty interventions.

Purpose of the Study:

• To introduce a methodology integrating feature selection with cross-validation for educational data analysis.
• To evaluate the stability and performance of different feature selection techniques.
• To identify key predictors of student performance in undergraduate science courses.

Main Methods:

• Applied a modified pipeline integrating feature selection and cross-validation to a dataset of 3225 students.
• Utilized 57 features, four DMMs, and four filter feature selection techniques.
• Employed Borda's method to rank features based on stability and predictive power.

Main Results:

• Correlation Attribute Evaluation (CAE) and Fisher's Scoring Algorithm (FSA) significantly improved Area Under the Curve (AUC) for logistic regression (LR) and elastic net regression (GLMNET).
• Relief Attribute Evaluation (RAE) demonstrated high instability and resulted in the poorest prediction performance.
• Borda's method identified grade point average, credit hours, and concept inventory performance as crucial predictors.

Conclusions:

• Integrating feature selection enhances the interpretability and actionability of predictive models in undergraduate education.
• CAE and FSA are recommended for stable and accurate feature selection in educational data mining.
• The methodology provides valuable insights for faculty interventions and stakeholder decision-making.