Enhancement of DNN-based multilabel classification by grouping labels based on data imbalance and label correlation

Multilabel classification (MLC) is a challenging task in real-world applications, such as project document classification which led us to conduct this research. In the past decade, deep neural networks (DNNs) have been explored in MLC due to their flexibility in dealing with annotated data. However, DNN-based MLC still suffers many problems. Two critical problems are data imbalance and label correlation. These two problems will become more prominent when a training dataset is limited and with a large label set. In this study, special neural network configurations were developed to enhance the performance of DNN-based MLC based on data imbalance and label correlation. The classification accuracy of minority labels and users-preferred labels was increased using customized label groups. The proposed method was evaluated using river restoration project documents and other fifteen datasets. The results show that the proposed method generally increases f1-score for minority labels up to 10%. Adding label dependence into label groups improves the f1-score of user-preferred majority labels up to 5%. The accuracy increase varies in different datasets.