Applying masked language model for transport mode choice behavior prediction

Transport mode choice behavior prediction is an important link in intelligent transportation and travel services, allowing people to choose the transport mode that suits them based on different recommendations. Many existing studies often posit that individuals primarily select their modes of transportation based on criteria such as the shortest travel distance, the least amount of time required, or the lowest cost. Contrary to these common assumptions, research shows a notable discrepancy between the anticipated travel behavior based on those criteria and the actual decision-making patterns observed, emphasizing the intricacy and variety in traveler preferences. To further explore the patterns of people’s transport mode choices, this study includes an in-depth examination of how trip distance impacts mode share in five major cities. This analysis brings to light the variability in mode share changes across different spatial environments, challenging the established utility theory for the mode choice model by accentuating the personalized nature of travel preferences in diverse contexts. To make more accurate and personalized predictions, this study further delves into the application of Masked Language Models (MLMs) for analyzing individual transport mode choice behaviors. It utilizes approximately two million trip records from major Chinese cities to train the model. In contrast to traditional methodologies that transform travel data into numerical values for utility calculation, this research employs a textual data approach. Travel information is reformatted into sentences that are processed by the MLM, which then converts these sentences into word vectors for input into a supervised learning model. The results indicate that this text-based approach maintains accuracy and broadens the scope of application by harnessing rich textual data, offering a distinct advantage over traditional numerical models.