RippleNet: Learning Causal Maritime Dynamics for Forecasting Warning-Induced Ripple Effects

Hongjun Wang, Zhiwen Zhang, Renhe Jiang, zipei fan, Xuan Song, Ryosuke Shibasaki

Keywords: Causal inference, Spatio-temporal forecasting, Maritime Warnings, Continuous-time dynamics, Ripple effects

Abstract: Maritime transportation networks, including cargo vessels, tankers, and passenger ships, are critical to global trade but remain highly vulnerable to disruptions such as extreme weather or security alerts. These events often trigger ripple effects, with cascading impacts extending far beyond the initial warning zones. Traditional spatio-temporal forecasting methods struggle to capture these dynamics due to their reliance on correlations rather than causal reasoning, particularly in maritime contexts. To address this challenge, we propose RippleNet, a novel causal spatio-temporal framework that explicitly models causal dependencies to predict port-to-port flow disruptions under warning-induced ripple effects. RippleNet comprises three key components: (i) a neural deconfounding module that employs causal adjustment techniques to disentangle genuine causal effects from spurious correlations, addressing confounding factors that arise when warnings simultaneously affect multiple maritime operational aspects, (ii) a continuous-time ODE module that simulates the propagation of disruptions across vessel networks, and (iii) LLM-generated warning vectors that quantify the multidimensional operational impacts of various warning types. Experiments on maritime flow datasets from East Asia and Northwest Europe show that RippleNet significantly outperforms state-of-the-art baselines under warning scenarios, while offering interpretable causal insights into heterogeneous vessel flow behavior.

Primary Area: learning on time series and dynamical systems

Submission Number: 10808