Enhancing partition distinction: A contrastive policy to recommendation unlearning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Growing privacy and data contamination concerns in recommendation systems necessitate effective recommendation unlearning methods.
• Existing approaches often neglect the balance between unlearning integrity, practicality, and efficiency, hindering real-world application.
• Prior studies suffer from imbalanced local and global collaborative information and data sparsity issues exacerbated by partitioning.

Purpose of the Study:

• To propose a novel data partitioning and submodel training strategy, PDCRU, for recommendation unlearning.
• To address the limitations of existing methods by balancing data partitioning and feature sparsity.
• To enhance both the learning and unlearning capabilities of recommendation systems.

Main Methods:

• Developed Partition Distinction with Contrastive Recommendation Unlearning (PDCRU) strategy.
• Extracted structural features as global collaborative information for samples.
• Introduced structural feature constraints based on sample similarity during partitioning.
• Leveraged contrastive learning for submodel training to enhance model embeddings.

Main Results:

• PDCRU demonstrates feasibility and consistent superiority over existing recommendation unlearning models.
• Achieved a 4.83% improvement in recommendation performance.
• Obtained a 4.64x enhancement in recommendation unlearning efficiency compared to baseline methods.

Conclusions:

• PDCRU effectively balances data partitioning and feature sparsity in recommendation unlearning.
• The proposed method offers significant improvements in both recommendation quality and unlearning efficiency.
• PDCRU presents a promising solution for privacy-preserving and efficient recommendation systems.