Instance-Level Difficulty: A Missing Perspective in Machine Unlearning
Machine unlearning research needs to focus on how difficult it is to remove individual data points. The study tested four factors—Tolerance of Performance Shift, Distance of Parameter Shift, Geometric Distance to Decision Boundary, and Number of Unlearning Steps—to quantify this difficulty. It found that while some factors correlate with success, a unified index reflecting instance-level variability is needed for better prediction.