ScholarWorks@숙명여자대학교

초록

Machine unlearning has emerged as a crucial concept to address data deletion requests arising from privacy protection and legal regulations, aiming to completely remove the influence of specific data from a trained model. Graph unlearning extends this concept to graph-structured data, where additional challenges arise from the need to account for relationships between nodes and edges. This study proposes a Node Embedding Unlearning method for Graph Convolutional Network (GCN)-based models that efficiently removes deleted nodes and their related information without performing full retraining. The proposed approach leverages the embeddings obtained during the initial training to recompute only the modified parts and utilizes sparse matrices containing information directly associated with the deleted nodes, thereby significantly reducing computational overhead. Experimental results demonstrate that the proposed method maintains comparable accuracy to full retraining while substantially reducing execution time. In particular, it proves effective in environments with low node deletion ratios, where real-time adaptation to graph changes is required.

키워드