Joint Deep Model with Multi-Level Attention and Hybrid-Prediction for Recommendation

Area of Science:

• Artificial Intelligence
• Machine Learning
• E-commerce Technology

Background:

• Recommender Systems (RS) are crucial in e-commerce, often using review text.
• Extracting informative features from vast review data and modeling user-item interactions remain challenging.
• Simultaneously capturing high- and low-order user-item interactions is an underexplored area.

Purpose of the Study:

• To develop an advanced recommender system that effectively utilizes review text.
• To address the challenge of modeling both low-order and high-order user-item interactions.
• To improve the accuracy and explainability of e-commerce recommendations.

Main Methods:

• A multi-level attention mechanism using Deep Neural Networks (DNN) was designed to identify useful reviews and significant words.
• A hybrid prediction structure integrating Factorization Machine (FM) for low-order interactions and DNN for high-order interactions was developed.
• The Multi-level Attentional and Hybrid-prediction-based Recommender (MAHR) model was built upon these components.

Main Results:

• Extensive experiments on Amazon and Yelp datasets demonstrated that MAHR outperforms existing state-of-the-art recommendation approaches in accuracy.
• Verification experiments and explainability studies, including attention module visualization, validated the model's effectiveness.
• The model successfully demonstrated the reasonability of its multi-level attention mechanism and hybrid prediction structure.

Conclusions:

• The proposed MAHR model offers a significant advancement in recommender systems for e-commerce.
• The multi-level attention and hybrid prediction approaches effectively address key challenges in review utilization and interaction modeling.
• MAHR provides more accurate and explainable recommendations, validating its innovative design.