This research article investigates the potential of Dual-encoder (DE) models in extreme multi-label classification (XMC) tasks. DE models, known for their success in dense retrieval tasks, are analyzed in the context of XMC tasks. The study reveals that with correct training, standard DE models can match or outperform state-of-the-art extreme classification methods. A differentiable top-k error-based loss function is proposed for optimizing Recall@k metrics. The PyTorch implementation of the model is included in the supplementary material.
Publication date: 16 Oct 2023
Project Page: https://arxiv.org/abs/2310.10636v1
Paper: https://arxiv.org/pdf/2310.10636
