Noise-Robust Preference Losses for Deep Regression Models
Keywords: Regression, Robustness, Alignment
Abstract: Deep regression models are widely employed for tasks such as pricing and forecasting. In industrial applications, it is common for analysts to adjust model outputs before they are deployed in commercial products. These adjustments, which we name "analyst influences", not only ensure the quality of the final products but also provide training data to improve model performance over time. However, due to the huge volumes of data, analyst influences can be applied broadly and can lack precision, hindering training effectiveness. To resolve the issue, we propose a novel framework Preference Learning from Analyst Influence which creates a weighted loss function that explicitly accounts for the relative quality levels of the training samples in comparison to model outputs. This approach effectively mitigates the impact of coarse training instances. Our extensive experiments on real-world data drawn from airline revenue management demonstrate that the proposed framework not only enhances pricing stability but also improves alignment with analyst influences compared to baselines.
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/2025/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: 8221
Loading