DGIST Researchers Develop AI 'Unlearning' Method to Preserve Recommendation Quality After Data Erasure

A research team at the Daegu Gyeongbuk Institute of Science and Technology (DGIST) has developed a new artificial intelligence technique designed to automatically identify and correct recommendation quality loss caused by the removal of user data. The method addresses a critical technical challenge facing digital platforms that must balance regulatory privacy mandates with maintaining personalized user experiences across automated systems.

Driven by global regulations establishing the "right to be forgotten," online services increasingly rely on a process known as "machine unlearning." The technology allows engineering teams to erase the influence of specific personal information from pre-trained machine learning models without undergoing the costly process of retraining an entire model from scratch. However, removing targeted data can cause collateral damage to the model's accuracy, diminishing recommendation quality for active users whose data remains in the system.

The DGIST researchers focused their study on graph neural network (GNN) architectures, which power modern recommendation systems by mapping relationships between users and items. Within these graph structures, certain users serve as essential structural links that bridge distinct user communities. When a user exercising privacy rights is removed from the network graph, those structural connections are severed, frequently leading to noticeable drops in recommendation accuracy for neighboring users.

To resolve this issue, the team created a mechanism named "Ada-Comp," short for Adaptive Compensation. As first reported by TechXplore, Ada-Comp operates by automatically detecting which user cohorts suffer the greatest decline in recommendation performance following a data deletion event. Once those vulnerable groups are identified, the system selectively delivers targeted additional training specifically for those segments, restoring recommendation accuracy while maintaining complete data erasure for the departing user.

The development shifts the traditional focus of machine unlearning research, which historically prioritized verifying that targeted personal data was thoroughly purged from a model. By analyzing post-deletion outcomes on the broader user base, the DGIST approach prevents unexpected performance drops from concentrating within specific user groups, ensuring broader fairness across platform demographics.

The study achieved international recognition at the ACM KDD Undergraduate Consortium (KDD-UC '26), held in Jeju in August. The paper was among just 28 research works selected globally for presentation at the conference, highlighting its relevance to ongoing discussions surrounding trustworthy artificial intelligence and privacy-focused software design.

The initiative highlights a successful interdisciplinary effort combining DGIST's undergraduate education with its professional research arms. Undergraduate researcher Eugene Jeong served as first author on the paper, collaborating with senior academic staff. Principal researcher Sang-Chul Lee from DGIST's Division of Nanotechnology acted as corresponding author, bridging student research with institutional research expertise.

"For trustworthy AI, it is important not only to delete personal data safely but also to ensure that the quality of service for other users does not deteriorate in the process," Lee said regarding the team's findings. He noted that researchers "expect this study to present a new direction for reducing performance disparities among user groups that may arise after machine unlearning."