Automatic Record Linkage for Structured Datasets Using Unsupervised Random Forests Model and Natural Break Thresholding

Abstract

Record linkage, which identifies the same entities across different datasets, is essential for ensuring data quality but is often manual and prone to errors. This study presents a two-step framework that automates threshold selection and generates training data using unsupervised machine learning and data-driven clustering. In the first step, an Unsupervised Random Forests (URF) model is used to compute similarity scores between record pairs without relying on labelled data. Four natural breaks algorithms, including Fisher (F), Jenks (J), Head/Tail (HT), and K-means (KM), are applied to these scores to determine upper and lower thresholds that separate matches from non-matches. These thresholds are then used to construct high-quality training data under “top” and “imbalanced” designs, supporting classifier training in the second step. The approach was evaluated on 13 datasets, including public benchmarks and synthetically corrupted data. Results showed that Random Forests (RF) with F or J breaks achieved the highest stability in numeric datasets, while Support Vector Machine (SVM) with J break performed best on string‑only data. Mixed datasets benefited from both RF and SVM, with performance depending on the proportion of textual versus numeric attributes. Compared with the Fellegi‑Sunter Expectation Maximisation and EpiLink baselines, the proposed approach achieved equal or superior F1 scores in most cases. Although performance declined in datasets with extreme corruption or poor imputation, the framework demonstrates that high‑quality linkage can be achieved without heavy feature engineering or deep embeddings. This highlights its value as a scalable, interpretable, and label-independent solution for structured datasets.