Assessing the Impact of Data Imbalance on the Predictive Performance of Machine Learning Models
Murali Mohana Krishna Dandu, Jaishree Jain, Santhosh Vijayabaskar, Punit Goel, Ashwini Shivarudra, Sachin Bhatt · 2024
Examining how data imbalance affects machine learning model prediction accuracy is the focus of this study. Here, the emphasis is squarely on a comprehensive architecture that integrates many classifiers with hybrid sampling methodologies. Machine learning has significantly transformed data analysis in several domains, with a special emphasis on healthcare, by improving the precision of diagnoses and decision-making processes. Imbalanced datasets, characterized by a large underrepresentation of one class relative to another, offer considerable issues. This often results in models that exhibit a bias towards the majority classes and perform poorly on the minority classes. To address this issue, our study utilizes a unique hybrid sampling technique that combines GASMOTE (Genetic Algorithm-based SMOTE) and SMOTE-PSO (Particle Swarm Optimization-enhanced SMOTE) with ENN under-sampling. This approach successfully addresses the problem of data imbalance. We use this framework to analyze health data obtained from Kaggle, combining it with XGBoost, Logistic Regression, Random Forest, Support Vector Machine, and a hybrid model of RF and XGBoost is not uncommon. A combination of these preprocessing procedures substantially improves the model’s performance, as shown by our study’s findings. No preprocessing was necessary with regard to hybrid RF+XGBoost model to achieve an accuracy of 0.97. Nevertheless, the model’s accuracy improved to 0.99 after preprocessing. Gains in accuracy were accompanied by enhancements in recall, precision, and F1-score. The study provides a standard structure to enhance the robustness and reliability with machine learning algorithms during circumstances with uneven data and highlights the pressing requirement for particular techniques to combat data imbalance.