Scholarlink

<p>Few-shot credit risk assessment on heterogeneous financial graphs is challenging due to scarce labeled defaults, shifting behavioral patterns, and highly uneven relational signals. Existing GNN-based models rely on static aggregation or single-level supervision, limiting their ability to adapt to evolving fraud behaviors or exploit unlabeled data effectively. We propose <b>AdaFuse</b>, a memory-augmented adaptive fusion framework designed for low-label and dynamic financial environments. AdaFuse introduces three key components: <i>Adaptive Multi-Relational Fusion</i>, which dynamically reweights transactional, social, and temporal relations through learnable gating and depth-aware aggregation; <i>Memory-Augmented Hierarchical Contrast</i>, which enhances representation discrimination via node-, subgraph-, and semantic-level contrast with a momentum-driven memory bank; <i>Meta-Learning–Enhanced Reconstruction</i>, which strengthens few-shot adaptability by integrating adjacency reconstruction and spectral regularization into a bi-level optimization scheme. These modules are trained jointly under a unified objective that balances contrastive alignment, structural preservation, and meta-level adaptability. Experiments on YelpChi, Amazon, and the large-scale FAIR-D dataset demonstrate that AdaFuse consistently surpasses state-of-the-art baselines in AUC, F1, and AP, achieving up to 5.6% absolute improvement. It further exhibits strong label efficiency, retaining over 92% of its full-supervision performance with only 10 labeled defaults per class. AdaFuse advances few-shot credit risk modeling by enabling dynamic relational adaptation and effective utilization of unlabeled financial graph structure, addressing key challenges in real-world, rapidly evolving credit ecosystems.</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

AdaFuse: memory-augmented adaptive fusion for few-shot fraud detection in heterogeneous graphs

  • 期刊论文

  • 发表日期：

    2026年1月24日

查看全文
  • Xuhui Wang,
  • Wayne Lin,
  • Jun Zhao,
  • Zhuang Liu

摘要

Few-shot credit risk assessment on heterogeneous financial graphs is challenging due to scarce labeled defaults, shifting behavioral patterns, and highly uneven relational signals. Existing GNN-based models rely on static aggregation or single-level supervision, limiting their ability to adapt to evolving fraud behaviors or exploit unlabeled data effectively. We propose AdaFuse, a memory-augmented adaptive fusion framework designed for low-label and dynamic financial environments. AdaFuse introduces three key components: Adaptive Multi-Relational Fusion, which dynamically reweights transactional, social, and temporal relations through learnable gating and depth-aware aggregation; Memory-Augmented Hierarchical Contrast, which enhances representation discrimination via node-, subgraph-, and semantic-level contrast with a momentum-driven memory bank; Meta-Learning–Enhanced Reconstruction, which strengthens few-shot adaptability by integrating adjacency reconstruction and spectral regularization into a bi-level optimization scheme. These modules are trained jointly under a unified objective that balances contrastive alignment, structural preservation, and meta-level adaptability. Experiments on YelpChi, Amazon, and the large-scale FAIR-D dataset demonstrate that AdaFuse consistently surpasses state-of-the-art baselines in AUC, F1, and AP, achieving up to 5.6% absolute improvement. It further exhibits strong label efficiency, retaining over 92% of its full-supervision performance with only 10 labeled defaults per class. AdaFuse advances few-shot credit risk modeling by enabling dynamic relational adaptation and effective utilization of unlabeled financial graph structure, addressing key challenges in real-world, rapidly evolving credit ecosystems.