Scholarlink

<p>Understanding and quantifying uncertainty in Graph Neural Networks (GNNs) is crucial for their deployment in real-world applications where data distributions often shift between training and inference prompting numerous proposed approaches. However, most of them primarily focus on uncertainty quantification under independent and identically distributed (i.i.d.) settings, overlooking the impact of distributional shifts, particularly conditional shift. This oversight can lead to unreliable predictions and miscalibrated uncertainty estimates, limiting GNN applicability in dynamic environments such as financial modeling, healthcare, and recommendation systems. To bridge this gap, we propose Conditional Shift Robust (CondSR), a novel, model-agnostic framework designed to enhance GNN performance under conditional shift. CondSR consists of a two-step approach: first, it regularizes the GNN training process using a loss function that explicitly minimizes conditional shift in latent representations, leading to improved model generalization. Second, it integrates conformal prediction to quantify uncertainty, ensuring that the model produces well-calibrated prediction sets with guaranteed coverage. Through extensive evaluations on benchmark graph datasets, we demonstrate that CondSR achieves predefined marginal coverage, improves GNN accuracy by up to 12% under conditional shift, and reduces prediction set size by up to 48%, significantly enhancing both reliability and efficiency. Our implementation is publicly available at <a href="https://github.com/Akanshaaga/CondSR.git">https://github.com/Akanshaaga/CondSR.git</a>, facilitating further research and adoption.</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

CondSR: Conditional Shift-Robust Conformal Prediction for Graph Neural Networks

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • S. Akansha

摘要

Understanding and quantifying uncertainty in Graph Neural Networks (GNNs) is crucial for their deployment in real-world applications where data distributions often shift between training and inference prompting numerous proposed approaches. However, most of them primarily focus on uncertainty quantification under independent and identically distributed (i.i.d.) settings, overlooking the impact of distributional shifts, particularly conditional shift. This oversight can lead to unreliable predictions and miscalibrated uncertainty estimates, limiting GNN applicability in dynamic environments such as financial modeling, healthcare, and recommendation systems. To bridge this gap, we propose Conditional Shift Robust (CondSR), a novel, model-agnostic framework designed to enhance GNN performance under conditional shift. CondSR consists of a two-step approach: first, it regularizes the GNN training process using a loss function that explicitly minimizes conditional shift in latent representations, leading to improved model generalization. Second, it integrates conformal prediction to quantify uncertainty, ensuring that the model produces well-calibrated prediction sets with guaranteed coverage. Through extensive evaluations on benchmark graph datasets, we demonstrate that CondSR achieves predefined marginal coverage, improves GNN accuracy by up to 12% under conditional shift, and reduces prediction set size by up to 48%, significantly enhancing both reliability and efficiency. Our implementation is publicly available at https://github.com/Akanshaaga/CondSR.git, facilitating further research and adoption.