Source-linked AI summary
DeepTravel: a Neural Network Based Travel Time Estimation Model with Auxiliary Supervision
Hanyuan Zhang, Hao Wu, Weiwei Sun, Baihua Zheng
TL;DR
Travel-time estimation must account for complex cross-segment dynamics while using the temporal labels available in trajectory data. DeepTravel is an end-to-end neural model with auxiliary supervision and multi-feature extraction, and experiments report advantages over existing competitors on real datasets.
Problem
Existing segment-based methods miss cross-segment factors, while heuristic sub-path methods do not fully leverage trajectory timestamps as supervised labels.
Method
DeepTravel directly predicts whole-path travel time using an end-to-end model with dual interval auxiliary supervision and spatial, temporal, driving-state, and traffic features.
Results
DeepTravel is significantly better than other end-to-end approaches across all reported metrics and outperforms competing methods on different datasets.
Takeaways & Limitations
The experiments demonstrate that DeepTravel can model different travel-time dynamics while addressing the separate-estimation and non-training-based limitations of prior approaches.
Takeaways & Limitations
Prior segment-based accuracy is constrained by speed estimates affected by sampling rate and GPS error, while sub-path approaches remain heuristic rather than training-based.
Abstract
from arXiv · showhide
Estimating the travel time of a path is of great importance to smart urban mobility. Existing approaches are either based on estimating the time cost of each road segment which are not able to capture many cross-segment complex factors, or designed heuristically in a non-learning-based way which fail to utilize the existing abundant temporal labels of the data, i.e., the time stamp of each trajectory point. In this paper, we leverage on new development of deep neural networks and propose a novel auxiliary supervision model, namely DeepTravel, that can automatically and effectively extract different features, as well as make full use of the temporal labels of the trajectory data. We have conducted comprehensive experiments on real datasets to demonstrate the out-performance of DeepTravel over existing approaches.
1 Introduction
Travel-time estimation is difficult because traffic, intersections, driving behavior, and historical variation interact across a whole path. DeepTravel addresses these limitations with end-to-end learning, auxiliary supervision from trajectory timestamps, and multi-feature extraction.
- 1 Introduction: Travel time is hard to estimate because traffic, crossroads, driving behavior, and historical path variation create complex dynamics.These dynamics make travel time indeterminate and difficult to estimate.
- 1 Introduction: Existing methods either estimate segments separately, missing intersection effects and accumulating errors, or estimate sub-paths heuristically without training-based learning.Segment-based methods also depend on speed estimates affected by sampling rate and GPS error.
- 1 Introduction: DeepTravel predicts the travel time of a whole path directly from historical trajectories through an end-to-end training-based model.The model is designed to learn directly from historical trajectory data rather than decomposing the path into separately estimated segments.
- 1 Introduction: DeepTravel introduces dual interval loss as auxiliary supervision to leverage the temporal labels provided by trajectory timestamps.The model also extracts spatial, temporal, driving-state, and traffic features for travel-time estimation.
- 1 Introduction: Experiments on two real datasets show that DeepTravel has an advantage over state-of-the-art competitors.The contribution passage reports comprehensive experiments and an overall advantage over competing approaches.
2 Related Work
Prior travel-time methods mainly estimate individual road segments or use historical sub-paths, while recent deep-learning work models trajectory sequences for related prediction tasks. These approaches motivate a learned path-level model for travel-time estimation.
- 2 Related Work: Segment-based methods estimate travel time for individual road segments, limiting their ability to represent interactions across a complete path.The related-work discussion places loop detectors, support vector regression, and stacked autoencoders in this category.
- 2 Related Work: Sub-path methods incorporate interactions between road segments by mining frequent patterns or using common sub-path travel times from historical trajectories.Examples include frequent trajectory-pattern mining and non-parametric estimation from shared sub-paths.
- 2 Related Work: Deep-learning methods have been applied to trajectory problems with recurrent networks that capture temporal or long-term dependencies.Prior applications include transportation-mode prediction and next-movement prediction.
3 Problem Definition
The problem is represented by mapping a sampled GPS trajectory onto a sequence of equal-sized grid cells. Empty cells remain in the sequence to preserve path continuity, enabling prediction of travel time for a query path.
- 3 Problem Definition: The road network is partitioned into N × N equal-sized grids, and a travel path is represented as the sequence of grids it traverses.The grid sequence is intended to capture the path’s movement through the road network when cells are sufficiently fine-grained.
- 3 Problem Definition: A trajectory is mapped to the corresponding grid path, with cells containing GPS points distinguished from cells without points.Figure 1 depicts occupied grids in blue and unoccupied grids in gray.
- 3 Problem Definition: Cells without GPS points are retained so the grid sequence preserves the continuity of the path.The representation therefore includes both observed and unobserved grid cells along the path.
4 Solution
DEEPTRAVEL represents paths with spatial, temporal, driving-state, and traffic features, then predicts travel time with BiLSTM and dual interval auxiliary supervision. The model uses short- and long-term traffic history and intermediate trajectory timestamps to supervise both whole-path and interval estimates.
- 4.1 Feature Representation Layer: DEEPTRAVEL combines spatial and temporal embeddings, driving-state features, and short- and long-term traffic features for path travel-time prediction.Features are carried by grid cells and extracted in a dedicated feature representation layer.
- 4.1 Feature Representation Layer: Short-term traffic features summarize recent grid conditions with LSTM-based temporal modeling, neighbor-grid information, and per-grid speed, sample-count, and travel-time estimates.The short-term window covers the previous hour, partitioned into five-minute bins in the described experiment.
- 4.1 Feature Representation Layer: Long-term traffic features extend the same extraction structure across days by using statistics for the same time in previous days.The described example uses information from the previous seven days.
- 4.2 Prediction Layer: BiLSTM combines feature representations across grid cells, while dual interval loss supervises forward intervals from the start and backward intervals to the destination.The loss activates supervision at grids containing trajectory information and uses intermediate timestamps in addition to whole-path travel time.
- 4.2 Prediction Layer: Summing hidden states for interval prediction encodes time additivity and lets the dual loss constrain both intermediate interval errors and whole-path travel-time estimation.The auxiliary supervision supplies additional monitoring information and improves the effectiveness of loss back-propagation.
5 Experiments
Experiments on real Porto and Shanghai trajectory datasets compare DEEPTRAVEL with existing and simple end-to-end baselines. DEEPTRAVEL performs best overall, while feature and auxiliary-supervision studies support its design choices.
- 5.1 Experiment Setting: The experiments use Porto and Shanghai trajectory datasets, with training, validation, and test splits set to 8:1:1.Porto contains data from 442 taxis, while Shanghai contains data from 13,650 taxis.
- 5.1 Experiment Setting: Simple grid-MLP, grid-CNN, and grid-LSTM models provide end-to-end baselines, but their comparison motivates DEEPTRAVEL’s richer feature extraction and model structure.The grid baselines use travel-length inputs and mean relative squared error.
- 5.2 Overall Evaluation: DEEPTRAVEL outperforms all competitors with significant advantages across the evaluated metrics and datasets.The comparison includes segment-based, sub-path-based, and end-to-end approaches.
- 5.3 Performance of DEEPTRAVEL: Combining spatial-temporal embeddings, traffic features, and driving-state features performs better than using individual feature groups or simpler historical-speed statistics.The full ST+Traf+DS configuration outperforms the evaluated feature variants.
- 5.3 Performance of DEEPTRAVEL: Auxiliary-supervision models perform better than models without auxiliary supervision, while BiLSTM captures grid correlations better than LSTM without auxiliary supervision.The loss-function study reports quantitative results in Table 4 and validation MAPE curves in Figure 4.
6 Conclusion
The paper concludes that DEEPTRAVEL addresses the limitations of segment-based and non-training-based path travel-time estimation. Its multi-feature extraction structure and dual interval loss support the reported superiority on real datasets.
- DEEPTRAVEL estimates whole-path travel time end to end, addressing the separate estimation of segment-based methods and the non-training drawback of sub-path methods.
- The model combines multiple feature types with a dual interval loss that incorporates additional supervisory information from trajectory data.
- Experiments on real datasets demonstrate DEEPTRAVEL’s superiority over existing approaches.