Ensemble deep random vector functional link neural net for imbalanced datasets

Abstract

In this project, the Ensemble Deep Random Vector Functional Link (edRVFL) network has been modified to enhance classification performance on any dataset with class imbalance. Imbalanced datasets is ubiquitous in numerous applications of machine learning and it is essential to develop models that could learn from these datasets and enhance its ability to classify minority classes in an accurate manner.

To achieve this, improvements along the machine learning pipeline has been introduced. Specifically, multiple sampling methods have been integrated into the pipeline and they will be used during training. Apart from pre-processing techniques, a novel cost function that addresses outliers and maximises classification accuracy for all classes has been implemented.

The newly proposed system has been tested on multiple imbalanced datasets that involve binary and multi-class classification. These experiments have demonstrated that this system has outperformed the generic edRVFL network for imbalanced datasets. Overall, both innovations have enhanced the model’s performance in classifying any imbalanced dataset.