Prompt-Tuning for Recommendation Unlearning

With the growing privacy concerns in recommender systems, the concept of recommendation unlearning is garnering increased focus. Existing methods, which primarily alter the recommender's parameters, tend to overlook the critical loss of valuable knowledge from the data not subject to unlearning. Inspired by the prompt learning in natural language processing, we proposed a prompt tuning method for recommendation unlearning. This method incorporates a teacher-student framework to facilitate forgetting and employs prompts in the user-item embedding space to theoretically match the efficacy of any form of prompting function. Experimental results demonstrate that our method outperforms existing recommendation unlearning baselines on standard evaluation metrics.