Recommendation model based on intention decomposition and heterogeneous information fusion

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Addressing timeliness and noise in user-item interaction data is crucial for effective recommendation systems.
• Existing methods struggle with capturing both short-term user intentions and long-term user/item features from diverse data sources.

Purpose of the Study:

• To propose a recommendation model, Intention Decomposition and Heterogeneous Information Fusion (IDHIF), that overcomes limitations in timeliness and data noise.
• To enhance user and item representation by effectively integrating information from user-item interactions, social networks, and knowledge graphs.

Main Methods:

• Decomposing user and item interaction intentions to extract short-term feature representations using Long Short-Term Memory (LSTM) and attention mechanisms.
• Employing heterogeneous information fusion to mine interactive, social, and content features from user-item interaction graphs, social graphs, and knowledge graphs.
• Combining short-term and long-term feature representations through splicing and multi-layer perceptrons to generate enriched user and item vectors.

Main Results:

• The IDHIF model demonstrated significant improvements over baseline models on the Last.FM and Movielens-1M datasets.
• Achieved increases in AUC (1.83-4.03%), F1 score (1.28-1.58%), and Recall@20 (2.90-3.96%).
• The model effectively enriches user and item vector representations, leading to enhanced recommendation performance.

Conclusions:

• The proposed IDHIF model offers a superior approach to modeling user and item features by integrating diverse information sources.
• The method effectively addresses the challenges of timeliness and noise, leading to improved recommendation efficiency and accuracy.
• This work contributes to the advancement of recommendation systems through sophisticated feature representation and information fusion techniques.