A Self-attention Graph Convolution Network Model for Estimating International Air Passenger Flows

Abstract: Understanding and forecasting the dynamics of international air passenger flows into the United States is a critical endeavor with profound implications for societal, economic, environmental, and public health considerations. In an era characterized by increasing global mobility, the ability to anticipate and manage the influx of travelers has become indispensable, impacting sectors ranging from tourism and transportation to hospitality and public policy. Moreover, it is vital to acknowledge the ecological toll exacted by air travel, with international flights accounting for approximately 102 grams of carbon dioxide and particulate matter emissions per kilometer traveled by each passenger. In 2019 alone, approximately 241 million international passengers arrived or departed from the United States, and projections indicate that this figure is poised to double over the next two decades, reaching nearly 500 million passengers annually. Despite the magnitude of international air travel and its substantial environmental footprint, the modeling of international passenger flows remains an underexplored domain. Existing forecasting approaches predominantly concentrate on individual airports, overlooking the intricacies of long-haul international travel and resulting in suboptimal predictions that fail to capture system-wide dependencies.
In response to these limitations, we present a novel and sophisticated Self-Attention Graph Convolutional Network (SAGCN) model designed to predict international travel patterns originating from approximately 50 foreign countries and terminating in the top 100 counties of the United States (Utilizing comprehensive quarterly DB1B international data spanning from 2017 to 2022). Our methodology is characterized by two pivotal components: (1) node embedding through a Graph Convolutional Network (GCN) and (2) the application of self-attention mechanisms to ascertain node-to-node connectivity, facilitating the extraction of valuable insights. The SAGCN model stands out for its robust performance, characterized by high generalization capabilities, and its ability to provide precise representations of inbound passenger flows. To substantiate the efficacy of our approach, we conduct comparative assessments against benchmark models, encompassing linear regression and multilayer perceptron neural network models.
In conclusion, our research contribution lies in the provision of a comprehensive, network-based framework for modeling and forecasting international air passenger flows, underpinning the sustainable and efficient planning of the global transportation system. Additionally, it is worth noting that our model's results have implications in the realm of Long-Distance Travel Demand Modeling (LDTM). Existing LDTM approaches often overlook the domestic component of international trips, thus, our model's results can be seamlessly integrated into LDTM efforts to accurately reflect long-distance travel patterns, accounting for both international and domestic aspects. As international travel continues to grow and exert pressure on our transportation systems and environmental sustainability, the utilization of advanced predictive tools becomes imperative. The comprehensive approach proposed through this effort provides actionable insights for industry stakeholders but also contributes to the broader discourse on sustainable travel and transportation planning, paving the way for a more efficient and environmentally conscious future.