Active Automated Machine Learning with Self-Training

Valentin Margraf, Julian Lienen, Eyke Hüllermeier, Marcel Wever

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.

Keywords: Active Learning, Semi-supervised Learning, Automated Machine Learning, Tabular Data

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

TL;DR: We develop the approach AutoActiveSelf-Labeling (AutoASL) which combines AutoML with self-training and active learning techniques to tackle real-world problems falling within the semi-supervised learning setting.

Abstract: Automated Machine Learning (AutoML) aims to automatically select and configure machine learning algorithms for optimal performance on given datasets. In real-world applications, training data oftentimes contain a large amount of unlabeled examples, whereas the amount of labeled examples is limited. However, AutoML tools have so far only focused on supervised learning, i.e., utilizing labeled data for training, leaving the valuable information provided by unlabeled data untapped. To address this limitation, we introduce our augmented AutoML system AutoActiveSelf-Labeling (AutoASL), which combines principles from self-training and active learning to effectively leverage unlabeled data during the training process. AutoASL iteratively self-labels previously unlabeled data instances, which is achieved through a powerful ensemble of AutoML and traditional ML algorithms, resulting in a substantial expansion of the labeled training data. We observe synergetic effects between the incorporated self-training and active learning components, leading to an improvement of the overall accuracy compared to state-of-the-art tools.

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: 3433