Scholarlink

<p>Marketing datasets are often highly sparse, especially in e-commerce environments where customers interact with only a limited subset of product categories. Although Product Hierarchy Modeling (PHM) provides a meaningful structure for representing category-level engagement, limited research has examined whether root-level sparse interaction patterns contain predictive signals for customer churn. Existing clustering approaches often rely on binarized categorical representations, such as Hamming-distance <i>K</i>-modes clustering, or on alternatives including hurdle-style clustering, embedding-based clustering, and Cosine-based hierarchical clustering, each involving trade-offs among computational tractability, interpretability, and preservation of engagement intensity. This study examines how the representation of sparse product-category interactions influences customer segmentation and churn prediction. Specifically, the study’s novelty lies in representing sparse categorical customer interactions as behaviorally meaningful PHM-based engagement patterns that preserve how customers allocate attention across product categories, with Cosine-based similarity serving as the intensity-preserving mechanism within this representation. The resulting clusters are incorporated into machine learning churn models alongside recency–frequency–monetary indicators. Analyses using B2B and B2C e-commerce datasets show stronger cluster cohesion, clearer separation, and more interpretable segments than binarized alternatives, while improving recall and reducing false negatives in churn prediction.</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

Clustering sparse product-interaction data: a cosine-versus hamming-distance approach to churn prediction

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Qiantianhong Wu,
  • Nora Sharkasi

摘要

Marketing datasets are often highly sparse, especially in e-commerce environments where customers interact with only a limited subset of product categories. Although Product Hierarchy Modeling (PHM) provides a meaningful structure for representing category-level engagement, limited research has examined whether root-level sparse interaction patterns contain predictive signals for customer churn. Existing clustering approaches often rely on binarized categorical representations, such as Hamming-distance K-modes clustering, or on alternatives including hurdle-style clustering, embedding-based clustering, and Cosine-based hierarchical clustering, each involving trade-offs among computational tractability, interpretability, and preservation of engagement intensity. This study examines how the representation of sparse product-category interactions influences customer segmentation and churn prediction. Specifically, the study’s novelty lies in representing sparse categorical customer interactions as behaviorally meaningful PHM-based engagement patterns that preserve how customers allocate attention across product categories, with Cosine-based similarity serving as the intensity-preserving mechanism within this representation. The resulting clusters are incorporated into machine learning churn models alongside recency–frequency–monetary indicators. Analyses using B2B and B2C e-commerce datasets show stronger cluster cohesion, clearer separation, and more interpretable segments than binarized alternatives, while improving recall and reducing false negatives in churn prediction.