Dominant preference decoupling and guided perturbed preference injection for cross-domain sequence recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Cross-domain sequential recommendation models user interactions across different domains.
• Existing methods often fail to account for dominant intra-domain preferences and the influence of perturbed preferences.

Purpose of the Study:

• To propose a novel model, Dominant Preference Decoupling and Guided Perturbed Preference Injection for Cross-Domain Sequence Recommendation (DP-CSR).
• To address limitations in current cross-domain recommendation systems by preserving dominant intra-domain preferences and leveraging perturbed cross-domain information.

Main Methods:

• DP-CSR employs multi-channel hypergraph learning and attention mechanisms to capture diverse intra-domain dominant preferences.
• It utilizes sequence encoders and a gating mechanism for dynamic injection of cross-domain perturbed preference information.
• Contrastive learning is used for preference decoupling and alignment between intra- and cross-domain perturbed preferences.

Main Results:

• DP-CSR demonstrates significant and consistent improvements over state-of-the-art methods.
• Experiments on six real-world datasets validate the model's effectiveness in cross-domain sequential recommendation.

Conclusions:

• The proposed DP-CSR model effectively enhances prediction adaptability by integrating dominant and perturbed preferences.
• The method successfully decouples and aligns preferences, leading to superior performance in cross-domain recommendation tasks.