Scholarlink

<p>This paper proposes a methodological framework for identifying economic clusters using unsupervised learning techniques applied to industry growth co-movements. Unlike conventional cluster identification approaches—based on product similarity, input–output linkages, or expert-defined industrial taxonomies—this method infers industry relatedness directly from observed dynamic behaviour. Using firm counts from the Registre des entreprises du Québec (REQ) between 1996 and 2024, we compute a correlation-based similarity matrix of annual industry growth rates and apply three clustering algorithms: hierarchical Ward clustering, K-means, and K-medoids. We evaluate each configuration using a validation metric, selecting the cluster definitions that maximize within-cluster coherence relative to between-cluster similarity. Our findings demonstrate that growth-based clustering reveals relationships that differ substantially from official industry classifications, identifying cross-sector linkages and dynamic co-vulnerabilities that static taxonomies do not capture. The paper concludes by discussing potential applications—including monitoring industrial resilience, detecting emerging clusters, and complementing traditional regional analysis—and outlines avenues for future methodological development.</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

Identifying industrial clusters through growth co-movement with unsupervised learning methods

  • 期刊论文

  • 发表日期：

    2026年1月30日

查看全文
  • Lucien Chaffa,
  • Thierry Warin

摘要

This paper proposes a methodological framework for identifying economic clusters using unsupervised learning techniques applied to industry growth co-movements. Unlike conventional cluster identification approaches—based on product similarity, input–output linkages, or expert-defined industrial taxonomies—this method infers industry relatedness directly from observed dynamic behaviour. Using firm counts from the Registre des entreprises du Québec (REQ) between 1996 and 2024, we compute a correlation-based similarity matrix of annual industry growth rates and apply three clustering algorithms: hierarchical Ward clustering, K-means, and K-medoids. We evaluate each configuration using a validation metric, selecting the cluster definitions that maximize within-cluster coherence relative to between-cluster similarity. Our findings demonstrate that growth-based clustering reveals relationships that differ substantially from official industry classifications, identifying cross-sector linkages and dynamic co-vulnerabilities that static taxonomies do not capture. The paper concludes by discussing potential applications—including monitoring industrial resilience, detecting emerging clusters, and complementing traditional regional analysis—and outlines avenues for future methodological development.