Sensitivity Analysis of Neural Network Ensembles to Data Quality Issues

The investigation of ensemble learning studies holds significant promise in machine learning research, offering the potential to improve performance and to address intricate data relationships. Despite its benefits, the performance of ensembles is often compromised by sub-optimal learners influenced by several data issues. This paper presents a sensitivity analysis of neural network-based homogeneous and heterogeneous ensembles to data quality issues. The study utilizes various classification datasets to systematically explore how variations in data quality, including skewed class distributions and outliers, impact predictive accuracy and generalization capabilities. The results indicate that the heterogeneous ensemble outperforms the homogeneous counterpart and a single neural network model across the skewed classes and outlier scenarios considered in the study. Overall, the findings contribute to a comprehensive understanding of the behaviour of neural network ensembles under diverse data conditions.