Problem: Educational institutions need to identify students who may be at risk of dropping out before the outcome becomes final. Student success is influenced by academic background, course selection, tuition status, debt, scholarship status, demographic factors, parental background, and semester-level performance. Goal: This project built a supervised learning workflow to predict whether a student would drop out, remain enrolled, or graduate, helping demonstrate how academic and administrative data can support early student-risk identification.

Initial Dataset: The original dataset contained 4,424 student records with 37 features. Final Modeling Dataset: After data cleaning, feature engineering, null handling, and outlier treatment, the final dataset used for model building contained 3,501 samples and 26 features. Target Variable: The classification target was the student outcome category: dropout, enrolled, or graduate.

Null Value Review: I identified missing values across columns such as Application order, Course, Day_class, Previous qualification, Previous qualification grade, Nationality, parental qualification, parental occupation, Admission grade, Displaced, Educational special needs, Debtor, Tuition fees up to date, Gender, Scholarship holder, and Target. Handling Strategy: Since the total number of missing values was small compared with the overall dataset size, I dropped rows with null values. For continuous variables such as Previous qualification grade and Admission grade, I reviewed their importance before final cleaning because they could contribute to student outcome prediction.

Outlier Analysis: I used box plots and value-count analysis to inspect unusual distributions and detect outliers across selected numerical and categorical features. IQR Method: I removed outliers using the Interquartile Range method where appropriate, reducing noise and improving the quality of the modeling dataset. Result: This preprocessing helped reduce the dataset from 4,424 records to 3,501 cleaner modeling samples.

Course Patterns: I observed that many students were enrolled in Health and Social courses such as nursing-related programs, while students in Basic Education-related courses showed a higher tendency to drop out. Financial Indicators: Students without debt and students with tuition fees up to date were more likely to complete and graduate. Scholarship-holder patterns were also reviewed because scholarship status showed differences in dropout behavior compared with other groups. Gender and Outcome Patterns: I analyzed dropout and graduation distributions across gender to understand whether outcome rates differed between male and female students.

Academic Progress Signals: There was a strong positive correlation between credited and approved curricular units in both semesters. The first-semester correlation was approximately 0.76, while the second-semester correlation increased to approximately 0.85, indicating that students credited for course units were highly likely to receive approval for those units. Demographic Patterns: Age showed a negative relationship with day-class enrollment, suggesting that older students were more likely to take evening classes and may also have different displacement patterns. Family Background: Parental education levels showed a positive correlation of approximately 0.57, and parental occupations showed a positive correlation around 0.49, suggesting partial alignment in family education and occupational background.

Categorical Transformation: I converted categorical features such as Nationality, Application mode, Course, and other grouped variables into model-ready encoded features. One-Hot Encoding: One-hot encoding was used to improve model compatibility and help categorical variables contribute more effectively to classification. Feature Reduction: After cleaning, grouping, encoding, and feature restructuring, the dataset was reduced from 37 original features to 26 final modeling features.

Models Tested: I trained and compared Linear Discriminant Analysis, Logistic Regression, K-Nearest Neighbors, Decision Tree, and pruned Decision Tree models. LDA Setup: I used the LSQR solver for LDA because it is numerically stable and suitable for high-dimensional data created through one-hot encoding. Logistic Regression Setup: I set max_iter to 1000 to give the model enough iterations to converge on the transformed dataset. KNN Setup: I selected the K value based on the value that produced the highest testing accuracy during model tuning. Decision Tree Setup: I used the Gini criterion because it is computationally efficient and suitable for larger structured datasets.

Training Accuracy: The Decision Tree model achieved the highest training accuracy, followed by Logistic Regression and the pruned Decision Tree model. Testing Accuracy: Logistic Regression achieved the highest testing accuracy, followed by LDA. F1-Score: Logistic Regression also achieved the highest F1-score, showing the best balance between precision and recall across the student outcome classes.

Best Model Selection: Logistic Regression was selected as the best model because it achieved the highest testing accuracy, strongest F1-score, and relatively strong training performance. LDA Insights: LDA also performed well, and confusion matrix analysis showed that it produced a high number of correct predictions compared with other interpretable models such as the pruned Decision Tree. Important Features: One-hot encoded features from Nationality, Application mode, and Course contributed strongly to LDA performance, showing that categorical restructuring helped improve class separability.

Outcome: The project successfully developed a supervised learning workflow to classify students into dropout, enrolled, or graduate categories. Model Result: Logistic Regression was selected as the best-performing model based on test accuracy and F1-score. Practical Value: The project shows how student demographic, academic, financial, and enrollment data can be transformed into a predictive workflow that supports early identification of students who may need additional academic or financial support.