Geo-Attention Network for Traffic Condition Prediction and Travel Time Estimation

Estimated time of arrival (ETA) is an important task in Intelligent Transportation Systems. Usually, the task involves a large amount of spatial-temporal data and is affected by different factors such as route distance, road capacity, traffic lights, and the real-time traffic condition. Real-time traffic conditions are highly uncertain and dynamic, which makes ETA challenging. For this reason, we propose an ETA model that incorporates the task of traffic condition prediction. Specifically, we introduce a Geo-Attention Network that combines a geo-location encoder and the geo-attentioned graph convolution to predict traffic conditions. Then, we use convolution network and recurrent neural network to capture the spatial and temporal correlations. Finally, we learn to estimate the arrival time and the traffic conditions simultaneously in a multi-task learning component. Extensive experiments have been carried out on the large-scale floating car data provided by GISCUP 2021, and excellent results have been achieved.