A COMPARATIVE STUDY OF DISTANCE METRICS AND NEIGHBOR SELECTION IN K-NEAREST NEIGHBOR FOR VOCATIONAL STUDENT PERFORMANCE CLASSIFICATION

Abstract

This study aims to evaluate parameter sensitivity in the K-Nearest Neighbor (KNN) algorithm, particularly the selection of distance metrics and k-values, for classifying academic performance in vocational education with heterogeneous and imbalanced data characteristics. The dataset consists of 750 first-year students from the Informatics Management program, including academic attributes (GPA, attendance, and core course grades) and demographic attributes (age, gender, educational background, and economic status). Data preprocessing involves data cleaning, one-hot encoding, Z-score normalization, and handling class imbalance using SMOTE. Model evaluation is conducted using K-Fold Cross Validation with accuracy, precision, recall, and macro-average F1-score as performance metrics. The results show that KNN performance is highly influenced by the combination of distance metrics and k-values. All metrics achieve accuracy above 84%, but differ in handling class imbalance. The Chebyshev metric (k = 10) provides the best balance with an F1-score of 0.6468, while the Minkowski metric (p = 3) achieves the highest recall of 0.7334. The Euclidean metric attains the highest accuracy of 0.8504 (k = 11), but tends to be biased toward the majority class. These findings indicate that optimizing KNN parameters should not rely solely on accuracy, but also consider balanced performance across classes. This study provides a practical evaluation framework for selecting KNN parameters to support more robust and fair academic prediction systems in vocational education data.