An Approach to Semantic-Aware Heterogeneous Network Embedding for Recommender Systems.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• Heterogeneous Information Networks (HINs) face challenges in recommendation systems due to data heterogeneity, particularly with unstructured user and item content.
• Existing HIN embedding methods struggle to effectively integrate diverse data types for improved recommendations.

Purpose of the Study:

• To propose a novel semantic-aware HIN embedding-based recommendation approach, SemHE4Rec, addressing data heterogeneity challenges.
• To develop a model that effectively learns representations from both structural and semantic information in HINs.

Main Methods:

• Introduced SemHE4Rec, a model employing two embedding techniques: co-occurrence representation learning (CoRL) using meta-path random walks and heterogeneous Skip-gram, and semantic-aware representation learning (SRL) for unstructured content.
• Integrated learned user and item representations with an extended matrix factorization (MF) process for joint optimization.

Main Results:

• SemHE4Rec demonstrated superior performance compared to state-of-the-art HIN embedding recommendation techniques on real-world datasets.
• The study confirmed that combining text-based (semantic) and co-occurrence-based (structural) representation learning significantly boosts recommendation accuracy.

Conclusions:

• The proposed SemHE4Rec model effectively addresses data heterogeneity in HINs by integrating structural and semantic information.
• Joint representation learning is crucial for enhancing the performance of HIN embedding-based recommendation systems.