Long Tail Classification Through Cost Sensitive Loss Functions
Keywords: Long Tail, Imbalanced Data, Cost-sensitive Loss
TL;DR: we introduce a novel Cost-Sensitive Loss (CSL) function designed to dynamically adjust class weights, and incorporate a reinforcement learning mechanism to optimize these adjustments.
Abstract: Class imbalance in the data introduces significant challenges in training machine models especially with long-tailed datasets. Specifically, it leads to biased models that overfit with respect to the dominant classes while under-performing on the minority classes. This, in turn, results in seemingly satisfactory yet biased overall results. Hence, the above biasing needs to be controlled such that the desired generalizability of the model is not entirely compromised. To that end, we introduce a novel Cost-Sensitive Loss (CSL) function designed to dynamically adjust class weights, and incorporate a reinforcement learning mechanism to optimize these adjustments. The proposed CSL function can be seamlessly integrated with existing loss functions, to enhance performance on imbalanced datasets, rendering them robust and scalable. We implemented the above CSL function in form of a framework which leverages reinforcement learning to optimally apply these adjustments over consecutive training epochs. Experimental Results on benchmark datasets demonstrate that our proposed approach significantly outperforms state-of-the-art methods. The results indicate that our approach can provide an optimal trade-off in the model accuracy and generalization with diverse kinds of imbalanced data.
Primary Area: datasets and benchmarks
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: 12484
Loading