The degradation of water quality, compounded by industrial activity and climatic stressors, complicates effective environmental monitoring and decision-making. Although conventional data-driven models offer predictive utility, they frequently function as opaque systems that prioritize statistical correlation over physical causality, leaving them vulnerable to distributional shifts. To transcend these limitations, this study proposes the Causal Mixture-of-Experts framework, which embeds causal structure discovery directly into a modular deep learning architecture. By utilizing the Nonlinear NOTEARS algorithm to derive transparent causal priors from observational data, the framework strictly constrains the expert system. Specifically, each expert is required to model a distinct data-generating subsystem to enforce mechanistic interpretability. Concurrently, systemic robustness is bolstered by a Causal Expert Modulation Module that integrates parent–child dependencies with a Drop-Expert regularization strategy to dynamically compensate for potential module failures. Rigorous evaluations across transnational, multi-source datasets from China, the United States, Canada, and the United Kingdom, conducted under strict comparisons with various advanced baselines, yield a classification accuracy exceeding 0.96. Notably, the model retains superior efficacy with scores remaining above 0.93 even under expert ablation, outperforming baselines such as TabKANet and LightGBM. Beyond predictive precision, the framework disentangles heterogeneous pollution drivers by identifying distinct mechanisms such as eutrophication dominance in China and organic pollution in Canada. This capability effectively bridges the divide between algorithmic modeling and accountable environmental stewardship.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号