Multi-attention graph neural networks for city-wide bus travel time estimation using limited data

An important factor that discourages patrons from using bus systems is the long and uncertain waiting times. Therefore, accurate bus travel time prediction is important to improve the serviceability of bus transport systems. Many researchers have proposed machine learning and deep learning-based models for bus travel time predictions. However, most of the existing models focus on predicting the travel times using complete data. Moreover, with the dramatically increasing population, bus systems also expand and upgrade their routes to provide improved coverage. Consequently, predicting the routes with sparse or no historical records becomes vital in this situation, and has not been well addressed in the literature. In particular, the challenges involved in this prediction include discovering routes with sparse records, discovering newly deployed routes, and finding the roads that need new routes. In order to address these, we propose a Multi-Attention Graph neural network for city-wide bus travel time estimation (TTE), especially for the routes with limited data, called MAGTTE. In particular, we first represent the bus network using a novel multi-view graph, which can automatically extract the stations and paths as nodes and weighted edges of bus graphs, respectively. Using inductive learning on dynamic graphs, we propose a multi-attention graph neural network with novel masks to capture the global and local spatial dependencies using limited data, and formulate a framework with LSTM and transformer layers to learn short and long-term temporal dependencies. Evaluation of our model on a real-world bus dataset from Xi'an, China demonstrates that the proposed model is superior compared to nine baselines, and robust to highly sparse data.