Domain-Agnostic Self-Training for Semi-Supervised Learning
Keywords: Self-training, Augmentation-free learning method, Semi-supervised learning, Tabular data
TL;DR: We propose DAST, a self-training framework that produces aligned data representations and reliable pseudo labels for domains without effective data augmentations.
Abstract: Self-training is a popular class of semi-supervised learning (SSL) methods which
can be viewed as iteratively assigning pseudo labels to unlabeled data for model
training. Despite its recent successes, most self-training approaches are domain-specific,
relying on the predefined data augmentation schemes in a particular
domain to generate reliable pseudo labels. In this paper, we propose a domain-agnostic
self-training framework named DAST, which is applicable to domains
where prior knowledge is not readily available. DAST consists of a contrastive
learning module along with a novel two-way pseudo label generation strategy.
Without the reliance of data augmentation, DAST performs supervised contrastive
learning with the pseudo labels generated from the classifier to learn aligned data
representations and produces the reliable pseudo labels for self-training based on
the learned representations. From an expectation maximization (EM) algorithm
perspective, we theoretically prove that representation learning and self-training in
DAST are mutually beneficial. Extensive experiments in various domains (tabular
data, graphs, and images.) verify that DAST not only significantly outperforms
other domain-agnostic self-training methods, but can also combine with effective
domain knowledge to further boost the performance.
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: 4927
Loading