1-Dimensional PCA Approach for Classifying Imbalanced Data Sets

Varayini Pankayatselvan, Derrick K Rollins, Sr.

“1-Dimensional PCA approach for classifying imbalanced data sets”

Major: Chemical Engineering

The goal of data mining is to uncover patterns within large sets of data as to reveal important information. It is a subfield of several disciplines, including statistics and computer science. Imbalanced binary data sets pose a challenge in extracting information to produce high classification results. Some methods developed to classify these sets involve multi-parametric classification algorithms such as SMOTE, neural network and random forest. These methods can produce relatively high classification results, however their methodology can also be very involved.

In bypassing the complexity, we have developed two powerfully simple one-dimensional classification methods producing high summary statistics. We implemented and modified a one-dimensional, PCA based approach, previously developed by Rollins et al., to separate imbalanced data sets on seven cases. Based on this PCA separation, we have developed two new classification techniques: mean method and spread method. In addition, we have also developed a feature reduction method in which we found an increase in classification results. We observed improved or similar results when comparing our data from the seven case studies with a recently published study on LVQ-SMOTE done by Kanazawa University in Ishikawa Japan.

The summary statistics were sensitivity, specificity, and g-mean. We note that only providing these statistics do not give enough information on the validity of the classification technique. We challenge that the actual values for true positives and false positives should be provided.