Comparison of Data Normalization Techniques on KNN Classification Performance for Pima Indians Diabetes Dataset
This study analyzes the comparison of data normalization techniques in the K-Nearest Neighbors (KNN) model for diabetes classification using the Pima Indians Diabetes dataset. The three normalization techniques evaluated are Min-Max Scaling, Z-Score Scaling, and Decimal Scaling. After preprocessing, such as handling missing values and removing duplicates, as well as feature selection using the Random Forest method, the features removed include SkinThickness, Insulin, Pregnancies, and BloodPressure. The evaluation was carried out using accuracy, precision, recall, F1-Score, specificity, and ROC AUC metrics. The results show that Min-Max Scaling provides a significant improvement in all metrics, with the highest accuracy of 0.8117 and ROC AUC of 0.8050. Z-Score Scaling provides good results, but not as good as Min-Max Scaling. Decimal Scaling shows the lowest performance. Statistical tests using Paired T-Test show significant differences between Min-Max Scaling and without normalization on all metrics (P-Value < 0.05), while Z-Score Scaling and Decimal Scaling are only significant on some metrics, with P-Values of 0.08363 and 0.43839 respectively for accuracy and ROC AUC. Overall, Min-Max Scaling proved to be the best normalization method for improving KNN performance in diabetes classification.
Standardized clinical MeSH and IEEE computational taxonomy keywords indexed for cross-database discoverability:
Empirical comparative analysis of Min-Max Scaling, Z-Score Scaling, and Decimal Scaling against unnormalized baselines to eliminate Euclidean distance distortion in KNN classification.
Applied Random Forest Gini Impurity feature selection strictly on training data (80:20 split) to prune 4 uninformative features (SkinThickness, Insulin, Pregnancies, BloodPressure) and prevent data leakage.
Comprehensive multi-metric benchmarking (Accuracy 81.17%, ROC AUC 0.8050, Precision 79.46%, Recall 79.70%, F1 79.58%, Specificity 84.85%) across neighbor variations K ∈ {1, 3, 5, 7, 9, 11}.
Rigorous Paired T-Test inferential statistical validation confirming that Min-Max Scaling produces statistically significant performance improvements (p < 0.05) across all six clinical evaluation metrics.
Pima Indians Diabetes dataset (768 patient records) inspected for datatypes, missing values (0), and duplicates (0), followed by an 80:20 Train-Test split.
Ensemble Gini Impurity ranking on training data isolates 4 high-impact predictors (Glucose, BMI, Age, DPF) while pruning 4 low-yield features.
Min-Max Scaling [0, 1], Z-Score Scaling (μ=0, σ=1), and Decimal Scaling applied to equalize feature variances and resolve distance bias.
Evaluated across K ∈ {1..11} with Euclidean metric; Paired T-Test confirms Min-Max superiority (p < 0.05 across all 6 diagnostic metrics).
Multi-format academic citation compiler with direct clipboard ingestion & reference manager file exports.
Pratama, Y. D., & Salam, A. (2025). Comparison of Data Normalization Techniques on KNN Classification Performance for Pima Indians Diabetes Dataset. Journal of Applied Informatics and Computing (JAIC), 9(3), 693-706. https://doi.org/10.30871/jaic.v9i3.9353