Context-Aware Concept Distillation：基于上下文感知的可解释洪涝预测框架（arXiv:2607.23237，2026-07-25提交，预印本）
论文主题为洪涝预测，不涉及三农、农业信息化、智慧农业、数字乡村或农业科技，不相关。
Levinkopf E、Morin E 与 Goldman CV 在 arXiv 提交预印本"Context-Aware Concept Distillation for Trustworthy Flood Prediction"（arXiv:2607.23237 [cs.LG]，2026-07-25提交）。研究面向公共安全高风险场景下深度学习"黑箱"阻碍信任与问责的难题，提出 Context-Aware Concept Distillation（CACD）框架，将 LSTM 蒸馏为可解释的水文学感知代理模型：通过无监督流程发现"水文学语言"，并用残差超网络（Residual Hypernetwork）根据静态流域特征动态调节这些概念。在全球 5,203 个流域上，模型中位 NSE 0.70，显著优于多层感知机等黑箱基线在未见过未来数据上的表现。
Abstract:Effective flood risk management relies on accurate forecasting, yet the "black box" nature of stateof-the-art Deep Learning models creates a barrier to trust and accountability in high-stakes public safety decisions. While existing Explainable AI (XAI) methods offer local attributions, they fail to provide the verifiable, operationally meaningful causal narratives required by disaster response authorities. To address this societal challenge, we propose Context-Aware Concept Distillation (CACD), a framework developed in collaboration with domain experts to distill opaque LSTMs into interpretable, hydrology-aware surrogate models. We introduce an unsupervised pipeline to discover a "Hydrological Language" and a Residual Hypernetwork that dynamically modulates these concepts based on static basin characteristics. Evaluated on 5,203 basins globally, our model achieves high fidelity (Median NSE 0.70), significantly outperforming black-box baselines (e.g., Multi Layer Perceptrons) on unseen future data. By demonstrating that human-interpretable concepts are sufficient to reconstruct flood dynamics, this work balances AI accuracy with the transparency required for responsible environmental decision-making.
Comments:to be published in IJCAI 2026 proceedings
Subjects:Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Cite as:arXiv:2607.23237 [cs.LG]
(or arXiv:2607.23237v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2607.23237
arXiv-issued DOI via DataCite (pending registration)
Submission history
From: Claudia Goldman [view email]
[v1] Sat, 25 Jul 2026 14:45:26 UTC (2,120 KB)