Improving Pre-Trained Models for Multi-Label Classification in Stack Overflow: A Comparison of Imbalanced Data Handling Methods

Tag classification is essential in Stack Overflow. Instead of combining through pages or replies of irrelevant information, users can easily and quickly pinpoint relevant posts and answers using tags. Since User-submitted posts can have multiple tags, classifying tags in Stack Overflow can be challenging. This results in an imbalance problem between labels in the whole labelset. Pretrained deep-learning models with small datasets can improve tag classification accuracy. Common multi-label resampling techniques with machine learning classifiers can also fix this issue. Still, few studies have explored which resampling technique can improve the performance of pre-trained deep models for predicting tags. To address this gap, we experimented to evaluate the effectiveness of ELECTRA, a powerful deep learning pre-trained model, with various multi-label resampling techniques in decreasing the imbalance that induces mislabeling in Stack Overflow's tagging posts. We compared seven resampling techniques, such as LP-ROS, ML-ROS, MLSMOTE, MLeNN, MLTL, ML-SOL, and REMEDIAL, to find the best method to mitigate the imbalance and improve tag prediction accuracy. Our results show that MLTL is the most effective selection to tackle the inequality in multi-label classification for our Stack Overflow data with deep learning scenarios. MLTL achieved 0.517, 0.804, 0.467, and 0.98 from the metrics Precision@l, Recall@5, F1-score@1, and AUC, respectively. Conversely, MLeNN gained only 0.323, 0.648, 0.277, and 0.95 from the same metrics.