Unlearning in Decision Tree Classifiers and Microcontroller Implementations

Machine unlearning allows the selective removal of information from trained models, addressing critical concerns such as bias, privacy, security, overfitting, and concept drift. In the realm of decision trees (DTs), unlearning a specific class, denoted as class-to-forget (CTF), traditionally involves entirely discarding the trained model and re-training a fresh model on the filtered data, which can be computationally expensive for large datasets and complex models. This paper explores an alternative approach to achieve unlearning: directly modifying the trained DT. Specifically, we focus on a method that removes nodes and leaves corresponding to the CTF. We delve into the design and implementation of this approach and analyze its computational effectiveness. We utilize set-theory-based mathematics and the sklearn library to design the unlearning methodology. We develop an embedded system using the STM32F303RE microcontroller Nucleo board and showcase the process of unlearning using an image dataset.