DA2-Unlearn: Dual-Adaptive Forget-Repair-Based Recommendation Unlearning

Modern recommender systems increasingly face recommendation unlearning requests, where specific user-item interactions must be revoked due to interest drift, accidental feedback, data poisoning, or privacy concerns. The forget-and-retain unlearning paradigm has gained increasing attention. However, existing approaches overlook two critical challenges: the varying difficulty of unlearning requests, and the delicate trade-off between effective forgetting and model utility preservation. In this work, we propose a dual-adaptive forget-repair-based recommendation unlearning framework, DA2-Unlearn. We first assess the intrinsic difficulty of each unlearning request and use it to modulate request-specific forgetting strength for effective request removal. We then perform a gradient-guided forgetting-first optimization, emphasizing early-stage forgetting and gradually shifting to restore recommendation utility, aligning the process with the principle of removing targeted interactions first to avoid meaningless repair. Extensive experiments on both real-world data with genuine revoking feedback and simulated datasets demonstrate that the proposed approach achieves significantly better unlearning effectiveness than existing methods, and meanwhile maintains comparable recommendation utility. Our code is available at https://gitlab.com/hcdou/da2unl.