Make Small Data Great Again: Learning from Partially Annotated Data via Policy Gradient for Multi-Label Classification Tasks

Zixia Jia, Junpeng Li, Shichuan Zhang, Anji Liu, Zilong Zheng

Keywords: Multi-label task, partially-annotated data, weakly supervised learning

TL;DR: We propose Partially Annotated reinforcement learning with a Policy Gradient algorithm (PAPG), a framework combining the exploration capabilities of reinforcement learning with the exploitation strengths of supervised learning.

Abstract:

Traditional supervised learning methods are heavily reliant on human-annotated datasets. However, obtaining comprehensive human annotations proves challenging in numerous tasks, especially multi-label tasks. Therefore, we investigate the understudied problem of partially annotated multi-label classification. This scenario involves learning from a multi-label dataset where only a subset of positive classes is annotated. This task encounters challenges associated with a scarcity of positive annotations and severe label imbalance. To overcome these challenges, we propose Partially Annotated reinforcement learning with a Policy Gradient algorithm (PAPG), a framework combining the exploration capabilities of reinforcement learning with the exploitation strengths of supervised learning. By introducing local and global rewards to address class imbalance issues and employing an iterative training strategy equipped with data enhancement, our framework showcases its effectiveness and superiority across diverse classification tasks.

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

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2768