Deep survival modelling for shared mobility

Bojan Kostic, Mathilde Pryds Loft, Filipe Rodrigues, Stanislav S. Borysov

TL;DR

This paper introduces deep survival modeling to predict car vacancy times in shared mobility systems, addressing the challenge of vehicle availability. By treating vacancy time as a survival event and employing a novel two-step approach with classification, the authors achieve a significant improvement in prediction accuracy, demonstrating the practical utility of these models for fleet optimization.

What Problem Does This Paper Address?

Shared mobility services, like car-sharing, are growing rapidly, aiming to reduce traffic congestion and environmental impact. However, a key operational challenge for fleet operators is ensuring vehicles are available where and when customers need them. This is particularly difficult in free-floating systems where cars can be left anywhere. To address this, operators often manually relocate vehicles to areas with higher demand. This paper tackles the problem of predicting how long a car will remain vacant at a given location, which is crucial for optimizing these relocation strategies and improving user satisfaction. Traditional methods often struggle with the "censored" nature of this data – we know a car was vacant until it was relocated, but not for how long it would have remained vacant.

Methods

The core of this paper is the application of survival analysis to model car vacancy times. Survival analysis is well-suited for "time-to-event" data, where the event (in this case, a car being picked up) might not be observed for all data points (due to relocation, which is called censoring).

The authors explore two main modeling approaches:

  • One-Step Approach: This is a standard approach where a survival model is directly applied to the entire dataset, predicting the probability distribution of vacancy times.
  • Two-Step Approach: This novel approach first uses a classification model to categorize cars into groups based on predicted vacancy time duration (e.g., short vs. long vacancy). Then, separate survival models are trained for each group. This aims to improve accuracy by handling different vacancy behaviors separately.

For survival modeling, they compare:

  • Cox Proportional Hazards (CPH) model: A traditional, linear survival model.
  • DeepSurv: A deep neural network-based survival model that can capture non-linear relationships.

For the classification step in the two-step approach, they compare:

  • Logistic Regression (LR)
  • Multi-layer Perceptron (MLP)
  • Gradient-Boosted Decision Trees (GBDT)

Additionally, they propose using Gaussian copulas to model spatial correlations between neighboring vehicles, allowing the status of one vehicle to influence the predicted survival functions of others.

Datasets

The primary dataset used in this research comes from SHARE NOW Copenhagen (formerly "DriveNow"), a free-floating electric car-sharing system. The data includes trip details such as vehicle ID, pick-up/drop-off times and locations, and battery levels, collected over approximately one year in 2017-2018.

Additional data sources used for feature engineering include:

  • OpenStreetMap: For road network information.
  • DTU Climate Station: For weather data (air temperature, wind speed, rainfall).
  • BikeRoads data: For road area types.
  • Danish National Travel Survey (TU): For socio-demographic information aggregated by zone.

The paper does not introduce a new benchmark dataset but uses the provided car-sharing data for training and evaluation.

Key Insights and Ideas

  • Vacancy time is a survival problem: Recognizing that car vacancy is a time-to-event phenomenon with censoring (due to relocation) is a crucial framing that allows for more appropriate modeling than standard regression.
  • Two-step approach significantly boosts accuracy: Classifying vacancy times before applying survival models leads to substantial improvements. This suggests that distinct behaviors for short vs. long vacancies are better captured by separate models.
  • Deep learning offers an edge: DeepSurv, the neural network model, generally outperforms the linear CPH model, indicating that non-linear relationships in the data are indeed beneficial for prediction.
  • Spatial correlation matters: Using Gaussian copulas to link neighboring vehicles' survival functions can refine predictions, reflecting how the status of one car can influence others.
  • Location is key, but not everything: While location (zone ID) is a significant predictor, incorporating more features and advanced models like DeepSurv yields better results.
  • Battery level is a critical operational factor: For electric cars, ensuring sufficient battery charge appears to lead to faster pick-ups, suggesting a clear operational strategy.

Key Results

  • The two-step approach achieved up to a 15% increase in C-index (a measure of prediction accuracy) compared to the one-step approach.
  • The DeepSurv model consistently outperformed the CPH model in the one-step approach, achieving a C-index of 0.683 compared to CPH's 0.658.
  • In the two-step approach, DeepSurv with a 1-hour threshold achieved the highest C-index, demonstrating the effectiveness of this combined strategy.
  • The GBDT classifier performed best in the classification step of the two-step approach, achieving high F1 scores across various thresholds.
  • An illustrative minimum cost flow problem demonstrated that the predicted survival times can be directly used for practical vehicle relocation optimization, showing a reasonable total travel distance for relocating cars.

Limitations and Future Work

  • Censoring data accuracy: The authors acknowledge that the current data doesn't precisely capture relocation times, leading to potential inaccuracies in censoring. More precise data on when cars are relocated would improve model performance.
  • Weather data influence: The impact of weather on survival times was less significant than expected, possibly due to the static nature of the weather data (measured only at drop-off) and its limited spatial resolution. Future work could incorporate time-dependent and spatially-resolved weather data.
  • Feature engineering: Reducing the dimensionality of features, especially the one-hot encoded zone IDs, could be beneficial. Exploring embeddings for zone characteristics (economic, demographic) is suggested.
  • Operational area expansion: The models could be extended to predict the attractiveness of new areas for expanding the shared mobility service.
  • More sophisticated correlation modeling: While Gaussian copulas were used, ideally, spatial correlation should be embedded directly into the multi-output survival model architecture for more precise time-dependent spatial correlation effects. This would, however, be computationally intensive.

Summary updated

Abstract

With an increased focus on minimizing traffic externalities in metropolitan areas, a growing interest in environmentally friendly and shared mobility systems has emerged, such as electric car-sharing systems. However, increasing demand and larger area coverage often make it difficult to keep cars available where and when customers need them. This problem can be alleviated by predicting for how long cars stay vacant at given pick-up/drop-off locations. To maximize their usage, it can be more beneficial to relocate the cars at certain periods to more desired locations. In this paper, we tackle the problem of predicting time-to-pickup for shared cars in a probabilistic way as a function of time by applying time-to-event modelling through survival analysis. Both statistical and deep neural network approaches to survival regression were investigated. The Cox proportional hazards model (CPH) is compared to the deep neural network model DeepSurv. To predict survival times, a two-step approach was formulated, where in the upper level a classification is used to classify cars into two groups based on idle time duration, whereas in the lower level for each given group time-to-event modelling is applied. DeepSurv method demonstrated a stronger fit compared to CPH. The two-step approach resulted in over 15% improvement in performance, comparing to the one-step approach, where no classification is used.
BibTeX
@inproceedings{kostic2021deep,
  author = {Bojan Kostic and Mathilde Pryds Loft and Filipe Rodrigues and Stanislav S. Borysov},
  title = {{Deep survival modelling for shared mobility}},
  booktitle = {Transportation},
  year = {2021},
  doi = {10.1016/j.trc.2021.103213},
}
Figure