Enhancing partition distinction: A contrastive policy to recommendation unlearning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Growing privacy and data contamination concerns in recommendation systems necessitate effective recommendation unlearning techniques.
• Existing methods often neglect the balance between unlearning integrity, practicality, and efficiency, hindering real-world application.
• Prior studies overemphasize local relationships and data partitioning exacerbates data sparsity, negatively impacting recommendation quality.

Purpose of the Study:

• To propose a novel data partitioning and submodel training strategy, Partition Distinction with Contrastive Recommendation Unlearning (PDCRU).
• To address the limitations of existing recommendation unlearning methods by balancing data partitioning and feature sparsity.
• To enhance recommendation quality and unlearning efficiency.

Main Methods:

• PDCRU extracts structural features as global collaborative information for samples.
• Structural feature constraints based on sample similarity are introduced during the partitioning process.
• Contrastive learning is employed for submodel training to enhance model embeddings with high-quality signals.

Main Results:

• PDCRU demonstrates feasibility and consistent superiority over existing recommendation unlearning models.
• The proposed method achieves a 4.83% improvement in recommendation performance.
• PDCRU offers a 4.64x enhancement in unlearning efficiency compared to baseline methods.

Conclusions:

• PDCRU effectively balances data partitioning and feature sparsity in recommendation unlearning.
• The method significantly improves both recommendation learning and unlearning performance.
• PDCRU presents a practical and efficient solution for recommendation unlearning challenges.