TRU: Targeted Reverse Update for Efficient Multimodal Recommendation Unlearning
Summary: arXiv:2604.02183v2 Announce Type: replace
Abstract: Multimodal recommendation systems (MRS) jointly model user-item interaction graphs and rich item content, but this tight coupling makes user data difficult to remove once learned. Approximate machine unlearning offers an efficient alternative to full retraining, yet existing methods for MRS mainly rely on a largely uniform reverse update across the model. We show that this assumption is fundamentally mismatched to modern MRS: deleted-data influence is not uniformly distributed, but concentrated unevenly across ranking behavior, modality branches, and network layers. This non-uniformity gives rise to three bottlenecks in MRS unlearning: target-item persistence in the collaborative graph, modality imbalance across feature branches, and layer-wise sensitivity in the parameter space. To address this mismatch, we propose targeted reverse update (TRU), a plug-and-play unlearning framework for MRS. Instead of applying a blind global reversal, TRU performs three coordinated interventions across the model hierarchy: a ranking fusion gate to suppress residual target-item influence in ranking, branch-wise modality scaling to preserve retained multimodal representations, and capacity-aware layer isolation to localize reverse updates to deletion-sensitive modules. Experiments across two representative backbones, three datasets, and three unlearning regimes show that TRU consistently achieves a better retain-forget trade-off than prior approximate baselines, while security audits further confirm deeper forgetting and behavior closer to a full retraining on the retained data.
Introduction
The growing complexity of Multimodal Recommendation Systems (MRS) has brought significant advancements in user experience and content personalization. However, the integration of user data within these systems presents challenges, particularly when it comes to removing or forgetting this data. Traditional methods often require full retraining, which can be inefficient and resource-intensive.
Challenges in Multimodal Recommendation Systems
The reliance on uniform reverse updates in unlearning methods poses several challenges:
- Target-Item Persistence: Items that have been marked for deletion can still influence the collaborative graph, leading to inaccuracies in recommendations.
- Modality Imbalance: Different modalities may have uneven representation within feature branches, complicating the unlearning process.
- Layer-wise Sensitivity: Certain layers within the model may be more sensitive to changes, requiring targeted updates to ensure effective unlearning.
Introducing Targeted Reverse Update (TRU)
To overcome these challenges, the TRU framework offers a novel approach. By focusing on targeted interventions, TRU enhances the efficiency of the unlearning process. The three key components of TRU include:
- Ranking Fusion Gate: This component minimizes the residual influence of deleted items within the ranking system.
- Branch-wise Modality Scaling: This ensures that retained multimodal representations are preserved, maintaining the integrity of the recommendation system.
- Capacity-aware Layer Isolation: This approach localizes reverse updates to specific modules, enhancing the model’s overall adaptability to changes in user data.
Experimental Results
Results from extensive experiments across various backbones, datasets, and unlearning regimes demonstrate that TRU significantly outperforms existing methods. The framework achieves a superior retain-forget trade-off, ensuring that the system can effectively forget outdated user data while retaining relevant information. Security audits conducted on the model further validate these findings, revealing that TRU’s performance approaches that of a full retraining process on retained data.
Conclusion
TRU represents a significant advancement in the field of multimodal recommendation systems, addressing the critical issue of data unlearning. By implementing targeted interventions, TRU not only improves efficiency but also enhances the overall user experience by ensuring that recommendations remain relevant and accurate. As MRS continues to evolve, frameworks like TRU will be essential in shaping the future of personalized content delivery.