Ethnicity-based name partitioning for author name disambiguation using supervised machine learning.

Area of Science:

• Bibliometrics
• Computer Science
• Information Science

Background:

• Author name disambiguation is crucial for accurate scholarly record-keeping.
• Existing methods face challenges, with certain ethnic name groups (e.g., East Asian) proving more difficult to disambiguate.

Purpose of the Study:

• To investigate the impact of ethnicity-based name partitioning on author name disambiguation performance.
• To determine if specialized models for ethnic name groups outperform general models.

Main Methods:

• Four machine learning algorithms were employed.
• Algorithms were trained and tested on both the entire dataset and on ethnicity-specific subsets.
• Performance was evaluated across different ethnic name groups and dataset configurations.

Main Results:

• Ethnicity-based partitioning substantially improved disambiguation performance across all tested ethnic name groups.
• Individual models trained on specific ethnic name groups were better suited for their respective data.
• Performance gains in correctly identifying matched name pairs exceeded losses in identifying non-matched pairs.
• Feature similarities, such as coauthor names, varied significantly across ethnic name groups.

Conclusions:

• Grouping author names by ethnicity before disambiguation offers a significant performance enhancement.
• Differences in feature similarities across ethnic groups suggest potential for developing ethnicity-specific feature weights.
• This approach is effective regardless of natural data distribution or controlled ambiguous name sizes.