Improving Multilabel Classification Model Performance in Imbalanced Datasets with Group-Level Undersampling
Danny Sebastián, Hindriyanto Dwi Purnomo, Irwan Sembiring · Engineering Technology & Applied Science Research · 2025
Multilabel classification presents unique challenges, particularly when dealing with imbalanced datasets. This study introduces a Group-Level Undersampling (GLU) technique designed to enhance the performance and efficiency of multilabel classification models. By converting multilabels into combined labels and categorizing them into group levels, the proposed method preserves the diversity of label combinations, which is crucial for accurate classification. Undersampling is particularly challenging in multilabel classification because removing a single instance can affect the combination of classes. This study utilized Google Maps reviews of tourist sites in Yogyakarta, Indonesia, manually labeled into 11 classes. The IndoBERTweet model was fine-tuned using the GLU generated dataset, and its performance was compared with models trained on randomly selected datasets. Experimental results demonstrate that the GLU method can maintain the variation of multilabel datasets by using combined labels and group levels. The experimental results show that the GLU dataset significantly improves model accuracy and F1 scores while reducing computational time and resources. The findings suggest that the proposed undersampling technique offers a robust solution to address imbalanced datasets in multilabel classification, paving the way for more efficient and accurate machine learning models. However, there are still weaknesses in the ratio calculation formula, allowing some datasets to be suboptimal at certain group levels. Future research should apply the GLU technique to various datasets and multilabel tasks to assess its generalizability. Investigating hybrid data-balancing methods could further enhance model performance and efficiency. A sensitivity analysis on ratioBaseLine values and refining the ratio calculation formula will improve the robustness of the GLU method.