Predictive Attributes in Machine Learning for University Academic Performance: A Feature Engineering Approach

This study focuses on identifying the most relevant predictive attributes that influence the academic performance of university students in Peru, the objective of this study is to identify the key predictive attributes that significantly influence the likelihood of predicting satisfactory academic performance. for this, advanced Machine Learning (ML) techniques and Educational Data Mining (EDM) were implemented, applying feature engineering (FE) methods such as redundancy elimination, significance analysis, handling of missing values, and the creation of new features. Algorithms like Random Forests (RF), among others, were utilized. The results demonstrated that FE significantly improves the performance of the models, especially in RF and Logistic Regression (LR), increasing the recall and the AUC ROC (0.94 RF and 0.93 LR), indicating a better ability to correctly identify students who will pass. The conclusion of the study indicates that characteristics are crucial to improving predictive models. Variables such as participation in tutoring, class attendance and use of academic resources are key. The study offers guidelines for implementing predictive systems that support at-risk students.