On Performance Comparison between Strong Machine Unlearning Algorithms for Logistic Regression Credit Assessment Models

The enactment of the Personal Data Protection (PDP) regulation in Indonesia requires financial service institutions to erase debtors' personal data upon request. However, this is challenging to achieve when such data is implicitly stored in a trained machine learning model. In order to address this issue, machine unlearning methods have been developed to erase the influence of training data on model weights. In this research, the performance of two strong machine unlearning algorithm implementations, ε-δ Certified Removal (CR) and Projective Residual Update (PRU), was compared on the logistic regression credit risk assessment models developed in this research. Machine unlearning was performed to delete 10% of the model's training data. Our results showed that ε-δ Certified Removal yielded a model with lower L²-distance, higher accuracy, and faster unlearning time compared to Projective Residual Update when machine unlearning was performed to delete less than 2% of the model's training data. Conversely, the opposite was observed when machine unlearning was performed to delete more or equal to 2% of the model's training data. Further research is required to explore the effect of larger training data sets with greater dimensionality on the performance of both algorithms.