Peeking the Impact of Points of Interests on Didi

Yonghong Tian, Zeyu Li, Zhiwei Xu, Xuying Meng, Bing ZhengarXiv 2018 · arXiv

Introduction

In on-demand transportation services like Didi, Uber, or Lyft, the "supply-demand gap" represents a critical operational challenge. This gap occurs when passenger demand significantly exceeds driver availability (e.g., during peak hours or bad weather) or vice versa. While external factors like weather and time are standard inputs for prediction models, the "internal DNA" of a neighborhood—its infrastructure, represented by Points of Interest (POIs)—is often underutilized.

This paper introduces PPCE-SDG (POI Principal Component Extraction Based on Supply-Demand Gap), a novel framework designed to identify which urban POIs—such as malls, schools, or parks—most significantly impact car-hailing demand.

Methodology: The PPCE-SDG Framework

The authors propose a three-stage pipeline to efficiently integrate POI data into demand estimation:

  1. Demand Clustering: Using K-Means++, city blocks are grouped based on similar supply-demand patterns over time, distinguishing between areas like residential zones and business districts.
  2. POI Aggregation: The distribution of 173 POI categories is calculated for each identified cluster.
  3. Principal Component Analysis (PCA): PCA is applied to these distributions to isolate the "heavy hitters"—the POI categories that explain the most variance in demand.

The selected significant POIs are then used as features in an XGBoost model to perform the final supply-demand gap estimation.

Data and Experimental Insights

The study utilizes a massive dataset from Didi Chuxing, covering 7 weeks of data from 58 blocks in Hangzhou, China, including over 11 million individual orders.

The "Power of Four" and Feature Bloat

A key finding is that a very small subset of POIs carries the majority of predictive power. The authors observed that the top four POI categories represent approximately 90% of the total impact on estimation accuracy. Furthermore, the research highlights the danger of "feature bloat." As shown in the accuracy curves, performance typically peaks and then declines as more POI categories are added, suggesting that irrelevant POIs introduce noise that confuses the model.

Figure 1: Estimation accuracy across different numbers of POIs for four distinct clusters. Accuracy tends to peak early, illustrating the diminishing returns of adding excessive features.

Performance Results

The PPCE-SDG method was evaluated against standard feature selection techniques. The inclusion of even a small number of POIs led to a significant jump in accuracy compared to models relying solely on external factors like weather.

Table 1: Comparison of estimation metrics with varying numbers of POIs. The inclusion of POI data (5 or 10 features) significantly improves average accuracy and reduces standard deviation compared to the baseline (0 POIs).

Beyond accuracy, PPCE-SDG achieved a Mean Absolute Error (MAE) of 3.41, outperforming other methods like XGBoost Feature Selection (3.47) and GBDT Feature Selection (3.62). The model also demonstrated higher stability, with lower standard deviation in its predictions across different time slices.

Conclusion

By treating POIs as the "genetic internal factor" of urban demand, the PPCE-SDG framework provides a more nuanced and efficient approach to supply-demand estimation. The results suggest that "peeking" at a few critical infrastructure markers is far more effective than broad, noisy data inclusion, allowing service providers to better anticipate and bridge the gap in urban mobility.

Summary updated

Abstract

Recently, the online car-hailing service, Didi, has emerged as a leader in the sharing economy. Used by passengers and drivers extensive, it becomes increasingly important for the car-hailing service providers to minimize the waiting time of passengers and optimize the vehicle utilization, thus to improve the overall user experience. Therefore, the supply-demand estimation is an indispensable ingredient of an efficient online car-hailing service. To improve the accuracy of the estimation results, we analyze the implicit relationships between the points of Interest (POI) and the supply-demand gap in this paper. The different categories of POIs have positive or negative effects on the estimation, we propose a POI selection scheme and incorporate it into XGBoost [1] to achieve more accurate estimation results. Our experiment demonstrates our method provides more accurate estimation results and more stable estimation results than the existing methods.
BibTeX
@misc{tian2018peeking,
  author = {Yonghong Tian and Zeyu Li and Zhiwei Xu and Xuying Meng and Bing Zheng},
  title = {{Peeking the Impact of Points of Interests on Didi}},
  year = {2018},
  eprint = {1804.04176},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1804.04176},
}
Figure