Scholarlink

<p>Entity resolution (ER) in large-scale graph databases is a core challenge in big data integration, where massive, heterogeneous, and continuously evolving datasets must be reconciled efficiently. Existing ER methods typically capture isolated aspects of the problem-structural, semantic, or temporal-limiting their effectiveness in real-world, high-volume environments. This paper introduces <i>DeepIDDFS</i>, a scalable node embedding designed for ER in large and dynamic graphs. <i>DeepIDDFS</i> employs an iterative deepening depth-first search (IDDFS) to efficiently explore multi-hop neighborhoods, integrates BERT-based semantic embeddings to handle noisy and inconsistent attributes, and incorporates a time-aware aggregation mechanism that emphasizes recent interactions to better model evolving graph structures. Extensive experiments on widely ER benchmarks—including DBLP-Scholar, Amazon-Google, and others—demonstrate that <i>DeepIDDFS</i> achieves state-of-the-art performance, reaching an F-measure of 0.94 and AUCPR of 0.95, while scaling to one million nodes with near-linear runtime growth across multiple domains. Ablation studies confirm the necessity of each component: removing semantic features reduces AUCPR by up to 18% and F-measure by up tp 15%, removing structural features reduces AUCPR by up 5% and F-measure by up 6%, and removing temporal weighting leads to reductions of up to 2–3% in both metrics. These results confirm that <i>DeepIDDFS</i> provides a robust solution for entity resolution in heterogeneous, dynamic big data environments. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DeepIDDFS: efficient entity resolution with hybrid embeddings

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Nour Mekki,
  • Djamel Berrabah,
  • Abdelhamid Malki

摘要

Entity resolution (ER) in large-scale graph databases is a core challenge in big data integration, where massive, heterogeneous, and continuously evolving datasets must be reconciled efficiently. Existing ER methods typically capture isolated aspects of the problem-structural, semantic, or temporal-limiting their effectiveness in real-world, high-volume environments. This paper introduces DeepIDDFS, a scalable node embedding designed for ER in large and dynamic graphs. DeepIDDFS employs an iterative deepening depth-first search (IDDFS) to efficiently explore multi-hop neighborhoods, integrates BERT-based semantic embeddings to handle noisy and inconsistent attributes, and incorporates a time-aware aggregation mechanism that emphasizes recent interactions to better model evolving graph structures. Extensive experiments on widely ER benchmarks—including DBLP-Scholar, Amazon-Google, and others—demonstrate that DeepIDDFS achieves state-of-the-art performance, reaching an F-measure of 0.94 and AUCPR of 0.95, while scaling to one million nodes with near-linear runtime growth across multiple domains. Ablation studies confirm the necessity of each component: removing semantic features reduces AUCPR by up to 18% and F-measure by up tp 15%, removing structural features reduces AUCPR by up 5% and F-measure by up 6%, and removing temporal weighting leads to reductions of up to 2–3% in both metrics. These results confirm that DeepIDDFS provides a robust solution for entity resolution in heterogeneous, dynamic big data environments.