Hierarchical deep learning for multi-label imbalanced text classification of economic literature
Shixun Lin, Flavius Frăsincar, Jasmijn Klinkhamer · Applied Soft Computing · 2025
With the vast amount of economic literature available in this day and age, efficient and accurate text classification becomes increasingly important. We propose an extended version of the Hierarchical Deep Learning for Text Classification (HDLTex) approach, called HDLTex++. HDLTex++ applies hierarchical learning using neural networks to classify documents and is adapted for the multi-label classification of class imbalanced data. We use HDLTex++ to assign to economic publications category labels from the Journal of Economic Literature classification system, which has a hierarchical tree structure with three levels. The performance of HDLTex++ is compared to two methods based on Support Vector Machines (SVMs), one where the class hierarchy is fully incorporated, and one where only the tertiary subcategories are taken into consideration. Performance is evaluated using the standard F1-score and a novel hierarchical F1-score that accounts for both class imbalance and class hierarchy. Our findings show that HDLTex++ is more effective in the prediction of primary category labels, compared to both SVM models, and in the prediction of secondary category labels, compared to the hierarchical SVM model. • We extend the hierarchical model HDLTex for multi-label, imbalanced classification. • We use this model to assign hierarchical category labels to economic publications. • We evaluate performance with both a standard F1 and a novel hierarchical F1 score. • We find the extended model performs well at predicting higher level labels.