Scholarlink

<p>Churn prediction, due to its imbalanced class distribution and high precision requirements on top-ranked instances, is particularly challenging among prediction tasks. Although in recent years, Graph Neural Networks (GNNs) have shown promise for churn prediction by leveraging user interactions, existing graph-based methodologies for handling class imbalance still encounter significant limitations regarding computational costs, structural distortion caused by synthetic or rewired edges, and incompatibility with inductive (near) real-time inference settings. To address these limitations, we propose a scalable subgraph training strategy that treats each node-centric subgraph as an independent training instance. This framework naturally supports inductive learning, avoids costly full-graph augmentation, and enables seamless integration with a wide range of class imbalance-handling techniques. To validate the efficiency of our approach, we conduct an exhaustive comparison against state-of-the-art full-graph methods across multiple GNN architectures on three large-scale, real-world telecommunications graphs consisting of millions of nodes. The results demonstrate that the subgraph strategy outperforms full-graph alternatives in terms of both AUC-PR and lift whilst offering better scalability and robustness.</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

When subgraphs outperform graphs: a scalable training strategy for churn prediction on large class-imbalanced networks

  • 期刊论文

  • 发表日期：

    2026年1月31日

查看全文
  • Yameng Guo,
  • Seppe vanden Broucke

摘要

Churn prediction, due to its imbalanced class distribution and high precision requirements on top-ranked instances, is particularly challenging among prediction tasks. Although in recent years, Graph Neural Networks (GNNs) have shown promise for churn prediction by leveraging user interactions, existing graph-based methodologies for handling class imbalance still encounter significant limitations regarding computational costs, structural distortion caused by synthetic or rewired edges, and incompatibility with inductive (near) real-time inference settings. To address these limitations, we propose a scalable subgraph training strategy that treats each node-centric subgraph as an independent training instance. This framework naturally supports inductive learning, avoids costly full-graph augmentation, and enables seamless integration with a wide range of class imbalance-handling techniques. To validate the efficiency of our approach, we conduct an exhaustive comparison against state-of-the-art full-graph methods across multiple GNN architectures on three large-scale, real-world telecommunications graphs consisting of millions of nodes. The results demonstrate that the subgraph strategy outperforms full-graph alternatives in terms of both AUC-PR and lift whilst offering better scalability and robustness.