From big data to decision support: a hybrid machine learning approach for strategic behavior of schedule recovery under service disruptions in global liner shipping

The reliability of liner shipping services is increasingly suffering from unexpected disruptions. Existing schedule recovery studies rely mainly on optimization models with simulated scenarios due to limited empirical data, which hinders both decision quality and practical applicability. Although a few studies have attempted breakthroughs through AIS data-driven approaches, they remain purely descriptive and are confined solely to port-skipping. This study proposes a novel data-driven framework to identify the mainstream recovery behaviors (including port-skipping, port-swapping, and speeding-up) from extensive AIS data. The resulting database covers the operational records of 3,559 container vessels across more than 700 ports worldwide between 2018 and 2023. Building on this, this study then develops a hybrid machine learning approach to predict the most suitable strategy. This developed model, named PHCUE, integrates a tailored hyperparameter optimization and dynamic-aware mechanism within the ensemble learning framework. Additionally, to mitigate the critical challenges inherent in shipping data, this study employs an advanced generative data balancing technique and utilizes a greedy feature selection method. Comprehensive evaluations demonstrate that the proposed model outperforms the existing approaches, achieving an overall accuracy of 0.84 and a macro-F1-score of 0.81. Furthermore, the interpretability analysis reveals that schedule recovery decisions are associated with a complex interplay of three core factors: port-centric economic value, direct operational costs and risks, and primary strategic preference. Ultimately, this work provides a robust predictive tool and actionable insights for various maritime stakeholders. It enables shipping companies to proactively plan voyages, while assisting port authorities in managing resources and policymakers in diagnosing resilience.