Interpretable Preference Elicitation: Aligning User Intent with Controllable Long-tailed Learning

HaiBin Wen, Shuangwang, Xucong Wang, YuHeng Yang, Pingxing Wang, Jiachang Zhan, Zhe Zhao, Pengkun Wang

Keywords: Long-tail learning

Abstract: Long-tailed recognition remains a significant challenge, where models often struggle with tail class performance and adaptability to diverse user preferences. While recent controllable paradigms leveraging hypernetworks allow numerical specification of head-tail trade-offs, defining these multi-dimensional preference vectors can be unintuitive for users. This paper introduces a novel framework that bridges this gap by enabling users to articulate their preferences through natural language. We propose a two-stage approach: first, optimal numerical preference vectors are identified for canonical distribution scenarios, and a rich corpus of corresponding textual descriptions is generated. Subsequently, a lightweight neural network learns to map sentence embeddings of these textual descriptions to the underlying 3D preference vectors controlling the expert ensemble. Our method significantly enhances the usability and interpretability of controllable long-tailed learning systems without compromising, and even slightly improving, their performance on benchmark datasets. This work facilitates more accessible and practical adaptation of long-tailed models to specific real-world requirements.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 24934