Evaluation measure for group-based record linkage

Area of Science:

• Data Science
• Information Science
• Computer Science

Background:

• Robust evaluation of record linkage techniques is crucial for accurate data integration.
• Existing measures (e.g., precision, recall) are inadequate for evaluating the quality of linked record groups.
• Group-based record linkage requires specialized methods for assessing clustering performance.

Purpose of the Study:

• To highlight the limitations of traditional evaluation metrics in group-based record linkage.
• To propose and validate a novel method for evaluating the quality of record clusters.
• To provide a more accurate and unambiguous assessment of linkage technique performance.

Main Methods:

• Developed a novel evaluation method to assess record allocation within predicted clusters against ground-truth data.
• Mapped predicted clusters to ground-truth clusters to categorize individual record assignments.
• Utilized seven distinct categories to reflect the accuracy of record grouping by linkage techniques.

Main Results:

• Empirically validated the proposed method using real-world data, demonstrating superior reflection of cluster quality compared to traditional measures.
• Showcased that traditional measures like precision and recall yield ambiguous results for group linkage.
• The proposed method provides unambiguous and detailed insights into linkage performance.

Conclusions:

• The novel evaluation method offers unambiguous results for group-based record linkage.
• The seven-category system provides detailed information on record prediction accuracy within clusters.
• This facilitates informed decisions on selecting appropriate record linkage techniques for specific applications.