Scholarlink

<p>We present SIGFN, a structural inference approach that leverages Generative Flow Networks to sample graph structures and node-wise parameters from a reward that couples data likelihood with graph regularization. Across curated benchmarks spanning 15–250 nodes and a single-trajectory regime, SIGFN often outperforms VAE-based and GFlowNet baselines, with stable relative gains despite moderate absolute AUROC. SIGFN integrates uncertain prior knowledge directly in the state space, improving robustness when traditional methods falter. On real-world traffic networks (PEMS03/04/07), SIGFN demonstrates applicability under noise and missing data. We clarify computational trade-offs under noisy priors and pruning, distinguish time vs memory complexity, and state the method’s limitations.</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

Harnessing Generative Flow Networks for Effective Structural Inference

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Aoran Wang,
  • Jun Pang

摘要

We present SIGFN, a structural inference approach that leverages Generative Flow Networks to sample graph structures and node-wise parameters from a reward that couples data likelihood with graph regularization. Across curated benchmarks spanning 15–250 nodes and a single-trajectory regime, SIGFN often outperforms VAE-based and GFlowNet baselines, with stable relative gains despite moderate absolute AUROC. SIGFN integrates uncertain prior knowledge directly in the state space, improving robustness when traditional methods falter. On real-world traffic networks (PEMS03/04/07), SIGFN demonstrates applicability under noise and missing data. We clarify computational trade-offs under noisy priors and pruning, distinguish time vs memory complexity, and state the method’s limitations.