Technology Category
Machine Learning: Multimodal LearningData Mining & Knowledge Management: Recommender SystemsNatural Language Processing: Language Grounding & Multi-modal NLP
Multi-Modal Recommendation Unlearning for Legal, Licensing, and Modality Constraints
🤖 AI Summary
Existing unimodal forgetting methods for multimodal recommendation systems (MMRS) fail to address targeted item deletion—required for legal compliance or license revocation—due to their reliance on matrix representations, data sharding, and costly aggregation, making them incompatible with multimodal graph structures.
Method: We propose the first Reverse Bayesian Personalized Ranking (Reverse BPR) objective tailored for MMRS, enabling fine-grained, shard-free, zero-training-cost item-level forgetting. Our approach jointly optimizes multimodal graph embeddings via dual objectives: preserving recommendations for retained items while explicitly suppressing interactions involving forgotten items.
Contribution/Results: The method achieves precise removal of target user–item interactions without degrading overall recommendation performance. Experiments on benchmark MMRS datasets show up to 49.85% improvement in recall over baseline forgetting methods and 1.3× faster inference than the Gold model, demonstrating both efficacy and efficiency.
Technology Category
Application Category
📝 Abstract
User data spread across multiple modalities has popularized multi-modal recommender systems (MMRS). They recommend diverse content such as products, social media posts, TikTok reels, etc., based on a user-item interaction graph. With rising data privacy demands, recent methods propose unlearning private user data from uni-modal recommender systems (RS). However, methods for unlearning item data related to outdated user preferences, revoked licenses, and legally requested removals are still largely unexplored.
Previous RS unlearning methods are unsuitable for MMRS due to the incompatibility of their matrix-based representation with the multi-modal user-item interaction graph. Moreover, their data partitioning step degrades performance on each shard due to poor data heterogeneity and requires costly performance aggregation across shards.
This paper introduces MMRecUn, the first approach known to us for unlearning in MMRS and unlearning item data. Given a trained RS model, MMRecUn employs a novel Reverse Bayesian Personalized Ranking (BPR) objective to enable the model to forget marked data. The reverse BPR attenuates the impact of user-item interactions within the forget set, while the forward BPR reinforces the significance of user-item interactions within the retain set.
Our experiments demonstrate that MMRecUn outperforms baseline methods across various unlearning requests when evaluated on benchmark MMRS datasets. MMRecUn achieves recall performance improvements of up to 49.85% compared to baseline methods and is up to 1.3× faster than the Gold model, which is trained on retain set from scratch. MMRecUn offers significant advantages, including superiority in removing target interactions, preserving retained interactions, and zero overhead costs compared to previous methods.
Problem
Research questions and friction points this paper is trying to address.
Unlearning outdated or legally restricted item data in multi-modal recommender systems
Addressing incompatibility of matrix-based unlearning with multi-modal interaction graphs
Improving performance and speed in unlearning while preserving retained interactions
Innovation
Methods, ideas, or system contributions that make the work stand out.
Uses Reverse BPR for forgetting marked data
Combines forward BPR to reinforce retain set
Improves recall by 49.85% over baselines
🔎 Similar Papers
No similar papers found.
