Effect of Label-Noise Filtering on Classification of Imbalanced Data Sets with SVM

The performance of classification models generated by supervised classifiers, such as Support Vector Machine (SVM), is significantly influenced by the quality of the training data set. However, in practical applications, we often encounter noisy, imbalanced data sets in which the size of one or more classes is smaller than desired. Oversampling, a technique that supplements original samples with synthetic ones, is effective in addressing these challenges to some extent. However, its usage comes with a trade-off. Since the original data set may contain noise, which can influence the generation of synthetic samples, oversampling can potentially amplify the noise level in the augmented data set. It is crucial to mitigate this possibility. One of the most effective methods to address this issue is employing suitable noise filtering as a pre-processing step. However, there is a significant risk of information loss, particularly when filtering the minority class. To provide insights into when it’s worthwhile to take this risk and which noise filtering methods to apply, we integrated various noise filters into several oversampling techniques before the synthetic sample generation step. We then assessed the impact on SVM model performance across data sets with varying noise levels. Our results indicate that utilizing the GMMNF filter as a pre-processing noise filter step is beneficial, particularly in scenarios with high noise levels. Remarkably, this approach does not induce any noticeable degradation even in noise-free cases.