Machine Unlearning for Random Forest via Method of Images

Hang Zhang, Kai Ming Ting

Abstract: Privacy law can now demand specific training samples, if requested from concerned parties, to be deleted from a trained model. Random forest, an effective and widely used machine learning algorithm, has been the model of study for various machine unlearning techniques. The current unlearning techniques of random forest involve additional processing before model training, so that fast unlearning of some samples can be achieved. However, no algorithm can achieve the unlearning of a trained random forest. This paper proposes a novel algorithm for unlearning a trained random forest. The algorithm employs the method of images to generate image samples of the samples that need to be forgotten and trains a small number of additional decision trees on these image samples. The proposed method, called MUMI, enables efficient unlearning of samples from a trained random forest. Our theorems and experiments show that MUMI achieves fast unlearning in a trained random forest with virtually no loss of model performance.

External IDs:dblp:conf/pkdd/ZhangT25