FragSel: Fragmented Selection for Noisy Label Regression
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: Noisy Labels, Regression, Mixture Models
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: To address the problem of regression with noisy labels, we propose the Fragmented Selection framework for selecting clean samples by Mixture of Neighboring Fragments and curate four benchmark datasets along with a novel metric, Error Residual Ratio.
Abstract: As with many other problems, real-world regression is plagued by the presence of noisy labels, an inevitable issue that demands our attention.
Fortunately, much real-world data often exhibits an intrinsic property of continuously ordered correlations between labels and features; where data points with similar labels are also represented with closely related features.
In response, we propose a novel approach named FragSel wherein we collectively model the regression data by transforming them into disjoint yet contrasting fragmentation pairs. This allows us to train more distinctive representations, enhancing our ability to tackle the issue of noisy labels.
Our FragSel framework subsequently leverages a mixture of neighboring fragments to discern noisy labels through neighbor agreement within both the prediction and representation spaces.
To underscore the effectiveness of our framework, we extensively perform experiments on four benchmark datasets of diverse domains, including age prediction, price prediction, and music production year estimation.
Our approach consistently outperforms thirteen state-of-the-art baselines, being robust against symmetric and random Gaussian label noise.
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: 3113
Loading