KHGCN: Knowledge-Enhanced Recommendation with Hierarchical Graph Capsule Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Knowledge graphs are increasingly integrated into recommendation systems to enrich item information and address data sparsity.
• Existing knowledge graph embedding and graph neural network methods can introduce noise in personalized recommendations.
• Effective representation learning from knowledge graphs is crucial for enhancing recommendation performance.

Purpose of the Study:

• To propose a novel model, the knowledge-enhanced hierarchical graph capsule network (KHGCN), for learning effective entity-relationship representations from knowledge graphs.
• To address the challenge of noise introduced by comprehensive knowledge graph information in personalized recommendation tasks.
• To improve recommendation accuracy by extracting hierarchical graph structures and filtering irrelevant information.

Main Methods:

• Developed the knowledge-enhanced hierarchical graph capsule network (KHGCN) model.
• Utilized entity disentangling to eliminate noisy entities and relationship representations.
• Employed an attentive mechanism for enhanced knowledge graph aggregation.
• Leveraged graph capsule networks to capture structured information between entities.

Main Results:

• The KHGCN model successfully extracts node embeddings while learning hierarchical graph structures.
• The model effectively removes noise from knowledge graph representations tailored for recommendations.
• Experimental validation on real-world datasets confirmed the model's effectiveness in improving recommendation performance.

Conclusions:

• The proposed KHGCN model offers an effective approach to leveraging knowledge graphs in recommendation systems.
• By learning hierarchical structures and filtering noise, KHGCN enhances personalized recommendations.
• The integration of graph capsule networks provides a more complete representation of entity relationships.