Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Feature Unlearning: Theoretical Foundations and Practical Applications with Shuffling
Authors: Yue Yang, Jinhao Li, Hao Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Across both tabular and image datasets, our empirical results show that our method not only effectively and efficiently removes the influence of designated features but also preserves the information content of the remaining features. |
| Researcher Affiliation | Collaboration | Yue Yang EMAIL Maincode; Monash University Jinhao Li EMAIL Monash University Hao Wang EMAIL Monash University |
| Pseudocode | Yes | Figure 1: An example of our feature unlearning algorithm. ...Our proposed feature unlearning consists of two key steps: random shuffling and fine-tuning. In the first step, the j-th feature is shuffled while its marginal probability distribution remains unchanged. ... The second step involves training the original model using the same loss function but with the shuffled dataset. |
| Open Source Code | Yes | The codes are available in the supplementary materials. |
| Open Datasets | Yes | This study employs six Open ML datasets [11]: Magic, Credit, and Cali for the single-feature unlearning setting... We further extend our unlearning method to an image classification task using the Celeb A [21]... |
| Dataset Splits | Yes | The test set, denoted as หDTest, is derived from หD using an 80:20 training-test split ratio. |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA V100 GPU. |
| Software Dependencies | Yes | All codes are implemented in Python 3.10 and Py Torch 1.12. |
| Experiment Setup | Yes | The batch size is set as 64 and the learning rate is set as 0.001. The optimizer used for neural network update is the Adam optimizer and the loss function is the cross-entropy loss function. ... During training, the original model fฮธ is trained for 1500 epochs. Models trained from scratch run for 2000 epochs to ensure convergence... For our approach and the two baselines, unlearning is performed over 1 to 1500 epochs... |