Imbalanced Dataset Handling in Forensic Software Features Extraction

Class imbalance often occurs in datasets that use the BIO scheme. This occurs during the process of extracting software features from online news. Information from online news only provides a few sentences with software features in the field of forensics. This causes an extreme gap between the minority class in sentences labeled B-FP and I-FP compared with the maj ority class O. This study compares methods for handling dataset imbalance, which can be a solution to this problem. This study uses the BIO scheme tag and methods for handling class imbalance using Random Oversampling, Random Undersampling, and Easy Data Augmentation. The results of this method were tested using the BiLSTM-MCRF model. This study shows that the Easy Data Augmentation method can increase the accuracy of the model being tested.