The Impacts of Dynamic Prices on Passenger Destination Prediction in Ride-on-Demand Services: A Conditional BiLSTM-Attention Model

Ride-on-Demand services (RoD) like Didi and Uber have become widely accepted among passengers, with dynamic pricing as their core feature. In passenger destination prediction task, it is common to consider the vehicle's prefix trajectory as the crucial feature. In this paper, we build upon existing studies by introducing dynamic pricing and proposing a conditional BiLSTM-attention model framework. The dynamic prices influence the training process of network's bidirectional information flow through the conditional recurrent mechanism. Experimental evaluation on a multi-source urban dataset shows that our model achieves a top-5 accuracy of 69.2 %, outperforming baseline models. Notably, considering dynamic prices in the early stages of the trajectory (e.g. the former 10% and 30% of the whole trajectory) improves the average accuracy by 48.0 % and 61.3 %, respectively. Our hope is that the study can provide insights for industries such as advertising and catering on how to better utilize dynamic prices to achieve their desired business objectives.