Multi-granularity preference enhancement with hierarchical feature extraction for session-based recommendations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Session-based recommendation systems predict user interactions using short-term behavior.
• Current methods struggle with the hierarchical nature of user preferences, limiting accuracy.
• User preferences have both coarse-grained and fine-grained characteristics.

Purpose of the Study:

• To propose a novel method, MPEHFE, for session-based recommendation.
• To accurately capture user preference signals by modeling item features at multiple granularities.
• To enhance personalization and recommendation accuracy.

Main Methods:

• MPEHFE models semantic item relationships at coarse and fine granularities.
• A differentiable architecture search enhances fine-grained preference modeling.
• Contrastive learning reinforces coarse-grained preferences by identifying noisy interactions.

Main Results:

• MPEHFE consistently outperformed state-of-the-art baselines on three benchmark datasets.
• Relative improvements of 3%-9% in Precision@20 (P@20) were achieved.
• Relative improvements of 11%-56% in Mean Reciprocal Rank@20 (MRR@20) were achieved.

Conclusions:

• MPEHFE effectively captures hierarchical user preferences in session-based recommendation.
• The proposed method offers significant improvements in recommendation accuracy.
• MPEHFE provides a robust framework for personalized session-based recommendations.