Subset Discovery for Entity Matching

Entity Matching (EM) is the task of identifying co-referent manifestations from multiple data sources. It has been a long-standing challenge in data integration, knowledge graph, and recommendation systems. In recent years, Machine Learning models are taking off in commercial EM pipelines. However, we observe that most challenging EM tasks have complex structures of local distributions, making ML model perform poorly when trained on the whole dataset. To mitigate this issue, we propose a rule-based subset discovery approach to identify multiple subsets. Our Subset Discovery Rule (SDR) are similar to conventional entity blocking rules, but aim at partitioning training data instead of filtering unmatched entity pair candidates. By training one EM model for each subset, the combined EM prediction can capture local data distributions, yielding accurate Entity Matching results. Naive SDR searching takes \(O(2^n)\) model retraining, while unsupervised clustering methods perform poorly. Therefore, we propose an \(O(n^2logn)\) Subset Discovery Rule Searching algorithm. We evaluate SDR on 11 EM benchmarks. The empirical results show that SDR is statistically comparable to SOTA EM methods. Compared with Random Forest (RF)-based EM, the F1-score of SDR-RF is 5.6% higher, and compared with Large Language Model (LLM)-based EM, the F1-score of SDR-LLM is 2.7% higher, showcasing its generalization property. Due to its model-agnostic design, SDR is an extensible plug-in for any machine learning-based EM model.