Towards Accurate Deep Learning Model Selection: A Calibrated Metric Approach

Yewen Fan, Nian Si, Xiangchen Song, Sathyamoorthy Rajendran, Kun Zhang

Keywords: Model Selection, Deep Learning, Calibration, Metric, Neural Networks, Deep Click-Through Rate Prediction Models, Stock Return Prediction Models

Abstract: The adoption of deep learning across various fields has been extensive, yet the methods for reliably evaluating the performance of deep learning pipelines remain underdeveloped. Typically, with the increased use of large datasets and complex models, the training process is run only once and the new modeling result is compared to previous benchmarks. This practice can lead to imprecise comparisons due to the variance in deep learning pipelines, which stems from the inherent randomness in the training process. Traditional solutions often require running the training process multiple times and are often infeasible in Deep Learning due to computational constraints. In this paper, we introduce a calibrated metric approach, designed to address this issue by reducing the variance present in its conventional counterpart. Consequently, this new metric improves the accuracy in detecting effective modeling improvements in the model selection stage. The efficacy of the new approach has been justified both theoretically and empirically.

Primary Area: other topics in machine learning (i.e., none of the above)

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 1174