Abstract
Class imbalance constitutes a significant challenge in text classification tasks, particularly due to the insufficient representation of minority-class instances. Traditional machine learning methods trained on imbalanced datasets tend to exhibit a bias toward majority classes, thereby limiting their ability to accurately classify minority-class samples. To address this issue, this study proposes a classification framework, termed Imb-LLMClass, to investigate the effectiveness of large language model-based classification approaches for imbalanced text data. Experimental results demonstrate that the proposed Imb-LLMClass framework achieved the highest performance among the compared methods when evaluated based on the average of three experimental runs, attaining a Macro F1-Score of 0.878 and a Balanced Accuracy of 0.870 across three datasets. Furthermore, the increase in the average minority-class recall to 0.834 indicates that the proposed approach not only improves overall classification performance but also enhances the learning of underrepresented classes. These findings indicate that, in imbalanced text classification tasks, evaluating large language models solely based on overall performance metrics may be insufficient and that their ability to effectively represent minority classes should also be taken into consideration.

This publication has 18 references indexed in Scilit: