Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Complementary Fusion of Deep Network and Tree Model for ETA Prediction

YuRui Huang, Jie Zhang, HengDa Bao, Yang Yang, Jian Yang

TL;DR

The paper tackles ETA estimation by combining a CNN-based neural network with gradient-boosting trees to leverage both sequence-aware patterns and structured features. It relies on rich feature engineering from Shenzhen traffic data, including time, spatial, and driver-history attributes, and trains eight neural models whose outputs are fused with tree-model predictions. The approach achieves top performance on SIGSPATIAL 2021 GISCUP benchmarks (A/B lists), demonstrating strong generalization and practical potential for transportation services. This fusion framework contributes a robust, data-driven ETA estimator that balances sequence modeling with powerful tree-based learning for real-world routing and pricing decisions.

Abstract

Estimated time of arrival (ETA) is a very important factor in the transportation system. It has attracted increasing attentions and has been widely used as a basic service in navigation systems and intelligent transportation systems. In this paper, we propose a novel solution to the ETA estimation problem, which is an ensemble on tree models and neural networks. We proved the accuracy and robustness of the solution on the A/B list and finally won first place in the SIGSPATIAL 2021 GISCUP competition.

Complementary Fusion of Deep Network and Tree Model for ETA Prediction

TL;DR

The paper tackles ETA estimation by combining a CNN-based neural network with gradient-boosting trees to leverage both sequence-aware patterns and structured features. It relies on rich feature engineering from Shenzhen traffic data, including time, spatial, and driver-history attributes, and trains eight neural models whose outputs are fused with tree-model predictions. The approach achieves top performance on SIGSPATIAL 2021 GISCUP benchmarks (A/B lists), demonstrating strong generalization and practical potential for transportation services. This fusion framework contributes a robust, data-driven ETA estimator that balances sequence modeling with powerful tree-based learning for real-world routing and pricing decisions.

Abstract

Estimated time of arrival (ETA) is a very important factor in the transportation system. It has attracted increasing attentions and has been widely used as a basic service in navigation systems and intelligent transportation systems. In this paper, we propose a novel solution to the ETA estimation problem, which is an ensemble on tree models and neural networks. We proved the accuracy and robustness of the solution on the A/B list and finally won first place in the SIGSPATIAL 2021 GISCUP competition.
Paper Structure (9 sections, 4 equations, 4 figures)

This paper contains 9 sections, 4 equations, 4 figures.

Table of Contents

  1. Introduction
  2. APPROACH
  3. Feature Engineering
  4. Modeling scheme
  5. Neural network structure
  6. Tree model scheme
  7. Ensemble
  8. EXPERIMENT
  9. CONCLUSION

Figures (4)

  • Figure 1: An example of estimated time of arrival (ETA), the white dot represents the starting position, the red point represents the destination position, the blue route represents the shortest travel route in time, and the other two gray lines represent other travel routes.
  • Figure 2: A path sequence containing only 4 links is used to illustrate the convolution structure. Compared with zhang2015sensitivity, we add three-dimensional link features after the embedding layer, and the other processes are the same as TextCNN.
  • Figure 3: The green box refers to sequence features and category characteristics, the yellow box refers to dense attributes, the convolution kernel uses one-dimensional convolution, and the number of circles represents the length of the convolution kernel.
  • Figure 4: The fusion structure of the tree model and the neural network. The fusion weight can be obtained by the performance of the verification set.