Learning with Temporal Label Noise

Sujay Nagaraj, Walter Gerych, Sana Tonekaboni, Anna Goldenberg, Berk Ustun, Thomas Hartvigsen

Primary Area: societal considerations including fairness, safety, privacy

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: Time Series, Label Noise, Healthcare, Crowdsourced Annotations, Mental Health, Subjective Annotations

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

TL;DR: We propose a novel problem setting in time series classification where label noise changes over time and provide robust solutions.

Abstract: Many sequential classification tasks are affected by label noise that changes over time. Such noise might arise from label quality improving, worsening, or periodically changing over time. In this work, we formalize the problem of label noise in sequential classification, where the labels are corrupted by a temporal, or time-dependent, noise function. We call this novel problem setting temporal label noise and develop a method to learn a sequential classifier that is robust to such noise. Our method can estimate the temporal label noise function directly from data, without a priori knowledge of the noise function. We first demonstrate the importance of modelling the temporal label noise function and how existing methods will consistently underperform. In experiments on both synthetic and real-world sequential classification tasks, we show that our algorithm leads to state-of-the-art performance in the presence of diverse temporal label noise functions.

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: 3956