Comparative Study of Undersampling Techniques in Text Classification

Imbalanced datasets pose critical challenges in text classification, often leading to biased predictions and reduced model performance. This study compares three undersampling techniques: Random UnderSampler, Near Miss, and Tomek Links, to evaluate their impact on five machine learning classifiers: Logistic Regression, Support Vector Machine, Random Forest, Naive Bayes, and Neural Network. The results demonstrate that the Random Forest classifier achieves the highest performance with precision of $\mathrm{9 7 \%}$, recall of $\mathrm{9 9 \%}$, and F1-score of $\mathrm{9 8 \%}$ across all methods, showcasing its robustness against data imbalance. The Near Miss technique improves Neural Network performance to a recall of $83 \%$ and F1-score of $\mathrm{7 1 \%}$, while Tomek Links further boosts Neural Network precision to $\mathrm{6 8 \%}$, recall to $\mathrm{7 2 \%}$, and F1-score to $\mathrm{7 0 \%}$. In contrast, simpler models like Naive Bayes exhibit limited improvements, with precision and recall scores around 47-51%. These findings underscore the importance of selecting appropriate undersampling strategies to enhance text classification outcomes and address data imbalance effectively.