Performance Comparison of Transformation Methods in Data Mining Classification Technique

Abstract

Transformation is a data preparation process for data mining. The main objective of this research was to compare five transformation methods in terms of classification accuracy that the transformed data provided. Those methods were the following: No transformation, Min-Max Normalization, Z-Score Standardization, Decimal Scaling, and Median Method. Three classification methods K-Nearest Neighbor, Artificial Neural Network, and Naïve Bayes were used to evaluate the transformation methods. Each of these datasets was divided into two groups at a ratio of 70:30. The first group was a training data set; the second group was a testing data set. The range of tested random seed parameter was from 10, 20, 30, 40 to 50. Six datasets were datasets of White Wine Quality, Pima Indians Diabetes, and Vertebral Column of which data were not much different and datasets of Indian Liver Patient, Working Hours, and Avocado of which data were much different.  All algorithms and procedures were implemented in R programming language. On 4 out of 6 tested datasets, transformation by Decimal Scaling and classification by K-Nearest Neighbor were the best combination, followed by transformation by Decimal Scaling and classification by Artificial Neural Network. Our findings may directly benefit those who are interested in efficiently mining some big data.

Keywords: Transformation, Min-Max Normalization, Z-Score Standardization, Decimal Scaling, Median Method, K-Nearest Neighbor, Artificial Neural Network, Naïve Bayes