Robust Sequential Recommendation With Decorrelation and Debiasing

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Transformer-based sequential recommendation (SR) algorithms have advanced significantly.
• Current SR methods face challenges including insufficient generalization, over-correlation, over-smoothing in self-attention, and popularity bias.

Purpose of the Study:

• To propose a robust SR model that overcomes existing limitations by decorrelating features and debiasing recommendations.
• To enhance the generalization and robustness of Transformer-based SR algorithms.

Main Methods:

• Implemented a hierarchical data augmentation strategy with new operations to enrich user sequences.
• Introduced an adaptive decorrelation module to mitigate over-correlation and over-smoothing.
• Utilized a disentanglement module to filter popularity bias and extract stable user preferences.
• Incorporated a regularization term to improve user preference representation and reduce overfitting.

Main Results:

• The proposed robust SR model with decorrelation and debiasing (RDD) achieved up to a 6.63% improvement over competitors.
• RDD demonstrated robustness against sparse interactions and cold-start problems.
• Experiments on benchmark datasets validated the model's effectiveness.

Conclusions:

• The RDD model effectively addresses key issues in Transformer-based SR, offering superior performance and robustness.
• The novel augmentation, decorrelation, and disentanglement techniques contribute to more reliable recommendation systems.