Author name disambiguation based on heterogeneous graph neural network

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• The increasing volume of academic publications and author overlap presents significant challenges for accurate author name disambiguation.
• Existing feature-based and connection-based clustering methods struggle with efficiency and accuracy in resolving author name ambiguities.

Purpose of the Study:

• To develop an advanced author name disambiguation method that overcomes the limitations of current approaches.
• To improve the accuracy and efficiency of assigning newly published papers to their correct authors.

Main Methods:

• Proposing a relational graph heterogeneous attention neural network to extract and represent semantic and relational information from papers.
• Incorporating multiple attention mechanisms to enhance the learning of diverse node and edge interactions within the graph.
• Modifying hierarchical clustering by using learned vector representations and automatically determining the optimal number of clusters (k-value).

Main Results:

• The proposed method achieved an average F1 score of 0.834 on the Aminer dataset.
• Demonstrated superior performance compared to existing mainstream author disambiguation methods.
• Ablation experiments confirmed the positive impact of multiple attention mechanisms on network performance.

Conclusions:

• The relational graph heterogeneous attention neural network offers a more accurate and efficient solution for author name disambiguation.
• The integration of graph topology and learned representations significantly improves clustering accuracy and automation.
• This approach effectively addresses the growing challenge of author name ambiguity in academic databases.