KG2Vec: A node2vec-based vectorization model for knowledge graph.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Traditional vectorization methods like Node2Vec, based on word2vec, are effective for homogeneous social networks but struggle with heterogeneous networks like knowledge graphs.
• Homogeneous network walk strategies in Node2Vec are unsuitable for knowledge graphs due to distinct node and edge features, leading to inaccurate predictions and vectorization.

Purpose of the Study:

• To propose KG2vec, a novel method for vector representation of heterogeneous networks, specifically knowledge graphs.
• To address the limitations of traditional methods in capturing full-text semantics and contextual relations within knowledge graphs.

Main Methods:

• Reconstructing the knowledge graph and applying a new random walk strategy tailored for heterogeneous networks.
• Developing two training models and optimization strategies to capture contextual environments between entities and relations.

Main Results:

• KG2vec effectively addresses insufficient context consideration in knowledge graph vectorization.
• The model shows improved performance, particularly in handling one-to-many relationships, outperforming traditional methods.
• Experimental results demonstrate higher accuracy compared to existing knowledge graph vectorization techniques.

Conclusions:

• KG2vec provides a semantically rich vector representation for heterogeneous networks by considering contextual relations.
• The proposed method offers a significant improvement over traditional approaches for knowledge graph vectorization, enhancing prediction accuracy and data quality.