An Empirical Study for Class Imbalance in Extreme Multi-label Text Classification
Sangwoo Han, Chan Ho Lim, Bonggeon Cha, Jongwuk Lee · 2021
Extreme multi-label text classification (XMTC) is the problem of finding the most relevant multi-labels from a text corpus with millions of labels. One of the key challenges in XMTC is that most labels appear only a few times, i.e., the class imbalance issue. To overcome the class imbalance problem, existing studies suggested various methods using different loss functions (i.e., focal loss function) and data augmentation (i.e., mix-up). In this paper, we investigate the effectiveness of two main approaches over the RNN-based and transformer-based deep XMTC models. In experimental results, we found that some improvement can be achieved when focal loss and mix-up are applied for deep XMTC models on various datasets.