ENHANCING RIDE SAFETY THROUGH REAL-TIME PREDICTION OF PASSENGER-DRIVER PAIR OUTCOMES

Abstract

In recent years, the pervasive adoption of online ride-hailing platforms, exemplified by industry leaders like Uber and Didi, has seamlessly integrated into urban transportation, significantly enhancing the convenience of our lives. The pivotal moment in this service occurs when a passenger and driver are matched through the platform, providing both parties with the autonomy to accept or cancel a ride with a simple click. The accuracy in predicting the success of this passenger-driver pairing, termed the matching success rate (MSR), emerges as a critical factor for ride-hailing platforms. This accuracy informs instant strategies such as order assignment, making the decision-making process inherently complex due to the dynamic nature of both driver and passenger interactions.
This challenge is distinct from traditional online advertising tasks, which typically involve predicting a user's response to an object, such as click-through rate predictions for advertisements. The uniqueness of the ride-hailing context requires a simultaneous consideration of dynamics from both the driver and passenger perspectives. Compounding this complexity is the significant data imbalance across different cities, posing challenges in training accurate models, particularly for smaller cities with limited available data. While sophisticated neural network architectures may enhance prediction accuracy in data-scarce scenarios, the intricacy of the design can hinder the model's ability to deliver timely predictions in a real-time production environment.
In response to these challenges, this paper introduces the Multi-View model (MV), a comprehensive approach that adeptly learns interactions among dynamic features encompassing passengers, drivers, trip orders, and contextual information to accurately predict the MSR of passenger-driver pairs. Addressing the data imbalance issue, the paper proposes the Knowledge Distillation framework (KD), which supplements the model's predictive capabilities for smaller cities by leveraging knowledge from cities with denser data. Additionally, KD generates a simplified model, ensuring efficient deployment without compromising accuracy.
The efficacy of the proposed solution is substantiated through extensive experiments conducted on real-world datasets from diverse cities. The results unequivocally demonstrate the superiority of the Multi-View model and Knowledge Distillation framework, validating their effectiveness in addressing the unique challenges posed by predicting the matching success rate in the context of online ride-hailing platforms.