A sequential recommendation method using contrastive learning and Wasserstein self-attention mechanism.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Transformer-based sequence encoders are effective for sequential recommendation.
• Existing dot product self-attention methods struggle with unpredictable user behavior and capturing collaborative transferability.
• Bayesian Personalized Ranking (BPR) loss lacks constraints, leading to suboptimal optimization.

Purpose of the Study:

• To propose a novel method using stochastic self-attention to address limitations in sequential recommendation.
• To enhance feature representation and collaborative transfer learning by incorporating uncertainty.
• To improve recommendation performance, especially for cold-start items.

Main Methods:

• Introduced uncertainty using elliptical Gaussian distribution with mean and covariance vectors.
• Employed Wasserstein self-attention to compute positional relationships and incorporate uncertainty.
• Generated high-quality positive samples using cloze and dropout mask mechanisms for multi-pair contrastive learning.
• Implemented a dynamic loss reweighting strategy to balance cloze and contrastive losses.

Main Results:

• The proposed model demonstrated superior performance compared to state-of-the-art methods.
• Significant improvements in Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG) were observed across multiple datasets (Beauty, Toys, ML-1M, ML-100M).
• The model showed particular effectiveness in handling cold-start items.

Conclusions:

• The novel stochastic self-attention method effectively addresses the unpredictability of user behavior in sequential recommendation.
• The integration of Wasserstein self-attention and Gaussian distributions enhances collaborative transfer learning and model robustness.
• The proposed approach offers a promising solution for improving recommendation accuracy and adaptability, especially in scenarios with limited user data.