Online Learning in Varying Feature Spaces with Informative Variation
Keywords: online learning, varying feature space, variation feature space, infomative message
TL;DR: Using informative variation to improve the performance of existing algorithms in varying feature space.
Abstract: Most conventional online learning literature implicitly assumed a static feature space, while in practice the feature space may vary over time with the emerging of new features and vanishing of outdated features, which is named as online learning with Varying Feature Space (VFS). There have been increasing attention that initiated the exploration into this novel online learning paradigm. However, none of them was aware of the potentially informative information embodied as presence / absence (i.e., variation in this paper) for each feature, which indicates that the existence of some features of this VFS can be correlated with the class labels. Such information can be potentially beneficial to predictive performance if properly used for the learning purpose. To this end, we formally formulate this specific learning scenario, namely Online learning in Varying Feature space with Informative Variation (OVFIV), and present a learning framework to address this problem. The essence of the framework aim for answering the following two questions: how to learn a model to capture the association of the existence of features with the class labels and how to incorporate such information into the prediction process in order to gain performance improvement. Theoretical analyses and empirical studies based on 17 datasets from diverse fields verify the validity of our proposed method.
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: 6699
Loading