Extreme multi-label (XML) classification involves assigning multiple labels to an instance from an extremely large
set of possible labels. Despite its significance, zero-shot learning within the context of XML classification remains relatively understudied. Zero-shot learning becomes pivotal when dealing with new labels not present during the training phase, a common occurrence in real-world applications. Existing approaches often resort to training zero-shot learning classifiers from scratch, which can be computationally expensive and may not fully exploit the knowledge embedded in pre-trained models. In this paper, we propose a novel approach to address this gap by introducing a method for transferring knowledge from a pre-trained XML classifier to enhance zero-shot learning capabilities. We present experimental results that demonstrate the potential of knowledge transfer from pre-trained XML classifiers as a promising avenue for advancing zero-shot learning in the challenging context of extreme multi-label classification.

Research Paper:


Source: Swiss Conference on Data Science (SDS)


author = {Ostapuk, Natalia and Dolomic, Ljiljana and Mermoud, Alain and Cudr\’e-Mauroux, Phillipe},
title = {Leveraging Pre-Trained Extreme Multi-Label Classifiers for Zero-Shot Learning},
booktitle = {Swiss Conference on Data Science (SDS)},
year = {2024},
url = {/assets/pdf/XML_ZeroShot_SDS2024.pdf}