Empirical Analysis of Loss Functions for Machine Unlearning Across Minority and Majority Classes
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Machine unlearning allows selective erasure of desired information or data points from models, especially in the event of privacy concerns, regulatory needs, or user requests. Machine unlearning under class imbalance is not explored. Works on Selective Synapse Dampening (SSD), SCalable Remembering and Unlearning unBound algorithm (SCRUB) and Bad Teacher approach assumes balanced class distributions, whereas real-world datasets often exhibit class distribution imbalance. The selection of the loss function has an impact on the effectiveness of unlearning. However, prior work has primarily employed the cross-entropy loss function. We provide the first systematic comparison of 6 loss functions across three SOTA unlearning methods on classimbalanced NLP datasets. Our primary focus is on selective data erasure for the majority and minority classes at once, while ensuring the functionality of a given retention set. Experimental evaluation is conducted on the Emotion and Hate Speech datasets, offering an extensive analysis of minority and majority class behavior changes before and after the unlearning process. The findings provide valuable insights into the inherent trade-offs between efficient data erasure and preservation of key model capabilities in natural language processing tasks.
Fields of Study & Subjects
Privacy-Preserving Technologies in Data
Machine Learning and Data Classification
Imbalanced Data Classification Techniques
Publication details
- Journal / source
- International Conference on Intelligent Systems for Communication, IoT and Security (ICISCoIS)
- Pages
- 1–7
- Publication date
- 8 January 2026
- Language
- EN