Secure Your Ride: Real-time Matching Success Rate Prediction for Passenger-Driver Pairs

Yuandong Wang, Hongzhi Yin, Lian Wu, Tong Chen, Chunyang Liu

TL;DR

This paper introduces a novel problem: predicting the success rate of matching passengers with drivers in ride-hailing services. They propose a "Multi-View" (MV) model to capture complex interactions between passengers, drivers, orders, and context, and a "Knowledge Distillation" (KD) framework to address data scarcity in smaller cities and enable efficient deployment.

What Problem Does This Paper Address?

Online ride-hailing platforms like Uber and Didi have become essential for urban transportation. A critical step in their operation is matching passengers with available drivers. However, after a match is made, either party can cancel, leading to inefficiencies and a poor user experience. Accurately predicting the "Matching Success Rate" (MSR) for a given passenger-driver pair is crucial for platforms to make optimal assignment decisions in real-time. This problem is more complex than traditional user response prediction (like click-through rates) because it involves the joint, dynamic decisions of two independent users (passenger and driver) influenced by a multitude of real-time factors. Furthermore, ride-hailing data is often geographically imbalanced, with major cities generating vast amounts of data while smaller cities have very little, making it difficult to train accurate models for the latter. Existing prediction methods are often too complex for real-time deployment or fail to address this data imbalance.

Methods

The authors propose a two-pronged approach:

  1. Multi-View (MV) Model: This model aims to capture the intricate interactions between various features. It breaks down features into four "views": passenger, driver, order, and context. These features are first embedded and then their interactions are modeled using three types of operations: inner product, element-wise product, and outer product. An attentive combination mechanism then fuses these interaction results to produce the final prediction. A key component of the MV model is the use of a Differentiable Neural Computer (DNC) with an external memory matrix. This allows the model to learn from temporal dynamics and, importantly, to store and transfer knowledge across different cities.
  2. Knowledge Distillation (KD) Framework: To address the data imbalance issue and enable efficient online deployment, the authors introduce a KD framework. This framework uses a complex "teacher" model (an enhanced MV model that leverages knowledge from data-rich cities) to guide a simpler, lightweight "student" model. The student model is trained to mimic the teacher's predictions, effectively transferring knowledge from high-resource cities to low-resource ones. This allows for accurate predictions in data-scarce environments while maintaining computational efficiency.

Datasets

The experiments were conducted on real-world datasets collected from six cities by Didi Chuxing: Beijing (BJ), Shanghai (SH), Shenzhen (SZ), Chengdu (CD), Dezhou (DZ), and Zhangzhou (ZZ). For the knowledge distillation experiments, BJ, SZ, SH, and CD were used as source cities (with abundant data), and DZ and ZZ were used as target cities (with sparse data).

Key Insights and Ideas

  • Bidirectional Decision-Making is Key: Unlike traditional user response prediction, MSR prediction requires understanding the joint decision-making process of both passengers and drivers, making it significantly more challenging.
  • Context is King (and Dynamic): Real-time factors like traffic conditions, supply-demand ratios, and even the time of day heavily influence matching success, necessitating models that can capture these dynamic contexts.
  • Data Imbalance is a Major Hurdle: The stark difference in data volume between major and smaller cities is a critical challenge for building universally effective models.
  • Memory Networks for Knowledge Transfer: The use of DNC with external memory is a novel way to not only capture temporal dynamics but also to store and transfer learned representations across different geographical (city) contexts.
  • Knowledge Distillation for Efficiency: KD provides a practical solution to deploy accurate models in resource-constrained environments by transferring learned intelligence from complex models to simpler ones.

Key Results

  • The proposed MV model consistently outperforms various feature interaction-based baselines (DeepFM, PNN, DIN, DCN, DSSM, Wide&Deep, DeepCrossing) on both AUC and RMSE metrics across all cities.
  • The DNC component with its external memory was found to be the most crucial for the MV model's performance, highlighting the importance of memory for capturing context and enabling knowledge transfer.
  • The KD framework significantly improves performance on target cities with sparse data compared to simple fine-tuning methods (Single-FT and Multi-FT).
  • The student model trained via KD achieves performance comparable to the teacher model, demonstrating its effectiveness in providing accurate predictions with much lower computational cost.
  • The MV model exhibits good scalability, with training time and inference time growing linearly with data size and showing efficient speedup with parallelism.

Limitations and Future Work

  • Hyperparameter Sensitivity: While generally robust, the MV model's performance can be sensitive to embedding size on target cities, potentially leading to overfitting with very large embedding sizes due to data scarcity.
  • Teacher Model Complexity: The teacher model in the KD framework is computationally heavier than the student model, though still more practical than some other complex models.
  • Future Work: The authors suggest that their approach can be generalized to other applications involving bipartite matching and decision-making in dynamic environments, such as friend-making websites or online marketplaces. Further exploration of more sophisticated knowledge transfer mechanisms and model optimization for even greater efficiency could also be valuable.

Summary updated

Abstract

In recent years, online ride-hailing platforms have become an indispensable part of urban transportation. After a passenger is matched up with a driver by the platform, both the passenger and the driver have the freedom to simply accept or cancel a ride with one click. Hence, accurately predicting whether a passenger-driver pair is a good match turns out to be crucial for ride-hailing platforms to devise instant order assignments. However, since the users of ride-hailing platforms consist of two parties, decision-making needs to simultaneously account for the dynamics from both the driver and the passenger sides. This makes it more challenging than traditional online advertising tasks. Moreover, the amount of available data is severely imbalanced across different cities, creating difficulties for training an accurate model for smaller cities with scarce data. Though a sophisticated neural network architecture can help improve the prediction accuracy under data scarcity, the overly complex design will impede the model's capacity of delivering timely predictions in a production environment. In the paper, to accurately predict the MSR of passenger-driver, we propose the Multi-View model (MV) which comprehensively learns the interactions among the dynamic features of the passenger, driver, trip order, as well as context. Regarding the data imbalance problem, we further design the Knowledge Distillation framework (KD) to supplement the model's predictive power for smaller cities using the knowledge from cities with denser data and also generate a simple model to support efficient deployment. Finally, we conduct extensive experiments on real-world datasets from several different cities, which demonstrates the superiority of our solution.
BibTeX
@article{wang2021secure,
  author = {Yuandong Wang and Hongzhi Yin and Lian Wu and Tong Chen and Chunyang Liu},
  title = {{Secure Your Ride: Real-time Matching Success Rate Prediction for Passenger-Driver Pairs}},
  journal = {IEEE Trans. Knowl. Data Eng.},
  year = {2021},
  url = {https://openreview.net/forum?id=A4lDsVjnae},
}
Figure