Classification for large number of variables with two imbalanced groups

In the presence of group imbalance and large number of variables problems, traditional classification algorithms tend to be biased towards the majority group. Several approaches
have been devoted to study such problems using linear and non-linear classification rules, but limited to group imbalance rather than the combination of both problems. This study proposed two algorithms of classification namely Algorithm 1 and Algorithm 2 which combine resampling, variable extraction, and classification procedure. The difference between the two algorithms is in terms of the order of resampling and variable extraction prior to the construction of linear discriminant analysis (LDA). Both simulated and real data sets were utilised to measure the performance of the proposed algorithms based on two evaluation indicators, sensitivity and specificity. Based on the findings, Algorithm 2 outperforms Algorithm 1 in classifying the minority group, while both proposed algorithms
perform equally well in classifying the majority group. Both proposed algorithms outperform the conventional LDA on principal components (PCA-LDA) in classifying the minority group. Also, this study has proven that the conventional PCA-LDA and conventional LDA are biased towards the majority group. Hence, both algorithms are suggested to be the alternatives for imbalanced classification with large number of
variables. Both algorithms are beneficial towards the practitioners of classification predictive modelling as well as statisticians in pattern recognition domain.