Automated Graph Attention Network for Heterogeneous Entity Resolution

Abstract

Entity Resolution (ER) aims to match records that refer to the same real-world entity. Homogeneous ER tasks assume that the attributes of two entities are consistent, allowing the use of rules to form attribute pairs or graph structures, which facilitate the model in capturing semantic connections. However, entities from different sources often have varying numbers and names of attributes, leading to heterogeneous ER tasks. In such tasks, graphs cannot be constructed using rules, and the use of graph structures in heterogeneous ER tasks significantly enhances matching effectiveness. We introduce AutoGAT, an end-to-end ER model that employs the attention mechanism to automatically mine semantic feature graphs and utilizes a hierarchical graph attention mechanism for matching. Testing on datasets demonstrates that AutoGAT excels in heterogeneous ER tasks, achieving up to a 9.1% improvement in F1 score compared to GPT-4o (few-shot), and up to a 6.4% improvement compared to Ditto.