Regulating Imbalanced Deep Models with User-Specified Metrics
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: Imbalance learning, Deep learning, Imbalance metrics, Classification
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Deep learning models implemented in real-world applications still face challenges from imbalanced data. Existing methods address the imbalance problem by balancing the models between the minority class and the majority class. However, practical applications may require an imbalanced optimization strategy that selectively unbalances the models and makes them more suitable for the applications than the balanced models. In this work, we first give a formal definition to accurately quantify the degree of imbalance of a model. Then, we propose a bias adjustment method that can efficiently optimize the model to a specified imbalance state according to application metrics or requirements so that this method has wide applicability. Finally, we introduce a training strategy that is advantageous to select the optimal representation parameters of the model during traditional training process. Extensive experiments verify the effectiveness and efficiency of our method, and compared with state-of-the-art algorithms, our method has significant improvement in different metrics including accuracy, F1 value and G-means.
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: 4525
Loading