PaperlayerView exact revision on arXiv

Source-linked AI summary

Contextualized Spatial-Temporal Network for Taxi Origin-Destination Demand Prediction

Lingbo Liu, Zhilin Qiu, Guanbin Li, Qing Wang, Wanli Ouyang, Liang Lin

arXiv:1905.06335v1cs.LGcs.AIstat.ML

TL;DR

The paper addresses the gap left by origin-only taxi demand prediction by forecasting demand between all region pairs for destination-aware preallocation. It proposes CSTN, which combines local spatial, temporal, meteorological, and global correlation contexts, and reports superior performance on both OD and origin-demand tasks.

  • Problem

    Existing taxi-demand methods mainly model departure locations and neglect passenger destinations, despite destinations being important for taxi preallocation.

  • Method

    CSTN integrates origin- and destination-view spatial modeling, ConvLSTM-based temporal and meteorological modeling, and similarity-weighted global regional correlations.

  • Results

    CSTN achieves an OD-MAPE of 24.93% and an O-MAPE of 12.92%, outperforming other state-of-the-art methods on both tasks.

  • Takeaways & Limitations

    The approach extends taxi demand prediction to interregional origin-destination forecasting and supports more destination-aware modeling of taxi demand.

Abstract

from arXiv · show

Taxi demand prediction has recently attracted increasing research interest due to its huge potential application in large-scale intelligent transportation systems. However, most of the previous methods only considered the taxi demand prediction in origin regions, but neglected the modeling of the specific situation of the destination passengers. We believe it is suboptimal to preallocate the taxi into each region based solely on the taxi origin demand. In this paper, we present a challenging and worth-exploring task, called taxi origin-destination demand prediction, which aims at predicting the taxi demand between all region pairs in a future time interval. Its main challenges come from how to effectively capture the diverse contextual information to learn the demand patterns. We address this problem with a novel Contextualized Spatial-Temporal Network (CSTN), which consists of three components for the modeling of local spatial context (LSC), temporal evolution context (TEC) and global correlation context (GCC) respectively. Firstly, an LSC module utilizes two convolution neural networks to learn the local spatial dependencies of taxi demand respectively from the origin view and the destination view. Secondly, a TEC module incorporates both the local spatial features of taxi demand and the meteorological information to a Convolutional Long Short-term Memory Network (ConvLSTM) for the analysis of taxi demand evolution. Finally, a GCC module is applied to model the correlation between all regions by computing a global correlation feature as a weighted sum of all regional features, with the weights being calculated as the similarity between the corresponding region pairs. Extensive experiments and evaluations on a large-scale dataset well demonstrate the superiority of our CSTN over other compared methods for taxi origin-destination demand prediction.

I. INTRODUCTION

The paper extends taxi demand prediction from origin-only requests to interregional origin-destination demand, motivated by destination-sensitive taxi preallocation. It proposes CSTN to integrate local spatial, temporal, meteorological, and global regional contexts.

  • Origin-only taxi demand models ignore passenger destinations, although destination distributions are critical for taxi preallocation systems.
  • Destination-aware preallocation addresses restrictions, rejected requests, inefficient trips, and passenger dissatisfaction caused by mismatched regional demand.
  • The proposed task predicts future taxi demand between any two regions rather than demand only at departure locations.
  • CSTN integrates local spatial context, temporal evolution context, and global correlation context through three dedicated modules.
  • The authors report extensive experiments on a large-scale benchmark showing CSTN outperforms existing state-of-the-art methods.

II. RELATED WORKS

The paper distinguishes citywide interregional taxi demand prediction from prior region-level demand forecasting and traffic-flow estimation at selected locations. It represents the city and origin-destination requests as tensors suitable for deep neural networks.

  • Prior taxi-demand methods forecast demand per region or at selected locations, whereas this work predicts interregional demand for taxi preallocation.
  • Traffic-flow estimation methods typically use counts from designed locations and struggle with high-dimensional origin-destination matrices or citywide taxi coverage.
  • The city is partitioned into H×W non-overlapping longitude-latitude grids, transforming taxi requests into matrix or tensor inputs.
  • Each taxi request contributes origin, destination, and timestamp information to the origin-destination demand representation.
  • The prediction task forecasts demand in a future interval from historical demand and meteorological data.

IV. THE PROPOSED METHOD

CSTN models taxi origin-destination demand through complementary local spatial, temporal, and global correlation contexts. Its local module explicitly processes origin and destination views before temporal modeling and feature fusion.

  • CSTN contains LSC, TEC, and GCC components for local spatial, temporal evolution, and global correlation context modeling.
  • A. Local Spatial Context Modeling: The LSC module uses origin and destination views to capture local spatial context from both directions of taxi demand.
  • A. Local Spatial Context Modeling: The DO matrix is generated by reshaping and transposing the OD matrix so each channel represents demand from one region to all destinations.
  • A. Local Spatial Context Modeling: A Two-View ConvNet processes the OD and DO matrices separately, then concatenates and fuses their features into a final local spatial feature.

B. Temporal Evolution Context Modeling

The TEC module models taxi-demand evolution from historical demand and changing meteorological conditions using ConvLSTM, preserving local spatial structure while accumulating sequential information.

  • B. Temporal Evolution Context Modeling: Meteorological conditions can affect future taxi demand, so TEC combines historical demand features with changing weather information.The paper gives sustained snowfall and reduced travel willingness as an example of weather-related demand variation.
  • B. Temporal Evolution Context Modeling: ConvLSTM is used instead of FC-LSTM because its convolutional connections preserve the structural locality of spatial features.Its memory cell also accumulates previous sequential information.
  • B. Temporal Evolution Context Modeling: After sequential ConvLSTM updates, the final hidden state h_n is passed through a convolutional layer to produce F_lt, encoding temporal evolution context.The ConvLSTM receives historical feature inputs sequentially and accumulates prior information in its memory cell.

C. Global Correlation Context Modeling

The GCC module captures demand correlations between distant but behaviorally similar regions, then combines global correlation features with local spatial-temporal features for OD-demand prediction.

  • C. Global Correlation Context Modeling: Global correlation context captures similar taxi-demand patterns among geographically distant regions with consistent attributes.Residential regions in different city areas may show similar morning-rush demand despite large physical separation.
  • C. Global Correlation Context Modeling: GCC computes each region’s global feature as a similarity-weighted sum of features from all regions.Regions with higher similarity receive greater relevance in the fused feature.
  • C. Global Correlation Context Modeling: The module embeds local spatial-temporal features into Fs, reshapes them into regional columns, and constructs a region-to-region similarity matrix.N equals H·W, and each column of Fs represents one region.
  • C. Global Correlation Context Modeling: GCC applies normalized similarity weights to aggregate all regional features into F_g, which is reshaped back to the spatial grid.The resulting global feature is fused with F_lt because global correlation lacks structural locality on its own.
  • C. Global Correlation Context Modeling: The fused feature F_ltg is fed to a linear regression layer to predict the future taxi origin-destination demand.The output is represented as X̂_t+1 with N filters in the prediction layer.

D. Implementation Details

The implementation specifies the principal channel and layer widths used in the LSC, TEC, and GCC components.

  • D. Implementation Details: The LSC module uses K=3, meaning each ConvNet contains three convolutional layers.This setting is explicitly specified for the CSTN implementation.
  • D. Implementation Details: TEC convolutional layers use 32 filters, and the local spatial-temporal feature F_lt has channel number C_lt=75.
  • D. Implementation Details: The GCC feature F_s uses channel number C_s=64.

V. EXPERIMENTS

The experiments establish a large-scale benchmark, define evaluation metrics, compare CSTN with state-of-the-art methods, and analyze the contribution of its components.

  • V. EXPERIMENTS: The study builds a large-scale benchmark for taxi origin-destination demand prediction.
  • V. EXPERIMENTS: It introduces evaluation metrics for the taxi origin-destination demand prediction task.
  • V. EXPERIMENTS: The experiments compare CSTN with several state-of-the-art methods and conduct component analysis for its modules.

A. NYC-TOD Dataset

NYC-TOD is a citywide taxi origin-destination benchmark built from 2014 Manhattan taxi trips and meteorological data, with temporal, spatial, and evaluation specifications.

  • Dataset construction: NYC-TOD is the first public benchmark for citywide taxi origin-destination demand prediction, combining taxi demand and New York City meteorological data.The last sixty days are used for testing, with earlier data used for training.
  • Spatial representation: Manhattan is divided into a 15 × 5 grid, with each grid representing a geographical region for origin-destination demand modeling.
  • Dataset construction: The dataset uses 132 million 2014 NYC yellow taxi trip records after excluding trips whose endpoints lie outside Manhattan.Each raw record contains timestamps and origin and destination coordinates.
  • Temporal and spatial distribution: Demand is aggregated into half-hour intervals, and most demand is concentrated in the city center and traffic hubs.The dataset contains more than ten million taxi requests per month.
  • Meteorological data: Meteorological inputs cover temperature, windchill, humidity, visibility, wind speed, precipitation, and weather conditions, encoded into a 29-dimensional vector.Weather conditions use one-hot encoding, while six numeric indicators are Min-Max normalized to [0,1].
  • Evaluation setup: Performance is evaluated with MAPE and RMSE after re-scaling predictions from the training range [−1,1] to normal demand values.Origin-destination and origin-only metrics are reported separately, while ground-truth values below 5 are filtered.

C. Comparison with the State-of-the-Art

CSTN is compared with historical, regression, boosting, multilayer, residual, and ConvLSTM baselines on whole-set, regional, weekly, and visualization analyses. It achieves the strongest reported performance across these comparisons, including consistent weekday and weekend superiority.

  • Whole testing set: CSTN achieves the lowest MAPE and RMSE on the whole NYC-TOD testing set for taxi origin-destination demand prediction.The comparison includes historical average, regression, XGBoost, MLP, ST-ResNet, and ConvLSTM methods.
  • High-demand regions: CSTN outperforms ConvLSTM by about 1% in MAPE for both origin-destination and origin-demand prediction on high-demand regions.The evaluation uses twenty regions covering about 70% of Manhattan’s taxi demand.
  • Different days: CSTN consistently outperforms competing methods in MAPE across all days of the week.Weekday and weekend averages are also compared, while HA-All and HA-Rec are excluded from the broader experiment because of poor performance.
  • Visualization: Figure 7 compares predicted and ground-truth interregional and origin demands across high-, middle-, and low-demand regions during one week.Predictions are shown in red and ground truth in blue.
  • Different days: Deep learning methods capture regular weekday patterns more effectively than less regular weekend demand patterns.

D. Component Analysis

Component analyses show that CSTN benefits from combining local spatial, temporal, global correlation, and meteorological contexts, while performance improvements saturate at a short input sequence.

  • Context components: The full LSC+TEC+GCC model reduces OD-MAPE to 27.27% and O-MAPE to 18.48%, yielding 2.5% relative improvement on average.The full model combines local spatial, temporal evolution, and global correlation context.
  • Local spatial context: Adding destination-view processing substantially improves the LSC model compared with using only the origin-view ConvNet.The comparison is reported in the Two-View ConvNet analysis.
  • Local and global context: Combining local and global features performs best, indicating that the two context types are complementary for taxi demand prediction.The local feature alone performs better than the global feature alone.
  • Meteorology: Meteorological inputs reduce O-MAPE from 20.03% to 19.41% for LSC+TEC and from 19.72% to 18.48% for CSTN.The experiments compare models with and without meteorological data.
  • Sequence length: Performance improves as the input sequence grows to five half-hour intervals, or 2.5 hours, after which longer sequences provide little obvious benefit.The model therefore uses sequence length n=5 in the experiments.

E. Further Discussion

Further discussion evaluates CSTN's runtime, long-term forecasting, and region-partition choices. CSTN runs in real time, while long-term accuracy declines over farther horizons and geographical partitioning slightly outperforms the ZCT-plus-land-use alternative.

  • Runtime Efficiency: 1.187 ms: CSTN predicts the next taxi origin-destination demand interval on an NVIDIA 1080 GPU.All compared methods achieve practical runtime efficiency, so runtime is not the bottleneck.
  • Long-Term Taxi Demand Prediction: L-CSTN extends CSTN with m decoding ConvLSTM units to forecast multiple future time intervals.The experiment uses five historical intervals and predicts six future intervals.
  • Long-Term Taxi Demand Prediction: 28.63%: L-CSTN's OD-MAPE for demand at t+2, while performance gradually drops as the predicted time interval increases.For demand at t+6, OD-MAPE increases to 30.86%.
  • Different Region Partition Manners: ZCT-plus-land-use-homogeneity divides Manhattan into 44 regions with relatively consistent land-use homogeneity.Historical origin-destination demand is organized as a 44×44 matrix, with each entry representing demand between an origin and destination region.
  • Different Region Partition Manners: The ZCT-plus-land-use-homogeneity partition achieves convincing OD-MAPE performance but remains slightly worse than geographical coordinates.The authors attribute this difference to weaker preservation of local spatial information in the 44×44 data organization.

VI. CONCLUSION

The conclusion frames taxi origin-destination prediction as necessary for more effective preallocation and presents CSTN as a unified contextual model. On 132 million NYC taxi trips, CSTN outperforms compared methods on both OD and origin-demand prediction, while region partition remains an open issue.

  • VI. CONCLUSION: Taxi origin-destination demand prediction estimates future demand between all region pairs, incorporating passenger destinations into taxi preallocation.The conclusion notes that city management rules and driver preferences can affect available supply between regions.
  • VI. CONCLUSION: CSTN integrates local spatial, temporal evolution, and global correlation contexts in one framework to predict demand for all region pairs.The model is presented as the paper's approach to learning taxi demand patterns from historical data.
  • VI. CONCLUSION: 24.93% OD-MAPE and 12.92% O-MAPE: CSTN outperforms other state-of-the-art methods on taxi OD and origin-demand prediction.The evaluation uses 132 million New York City taxi trip records.
  • VI. CONCLUSION: City region division remains an open problem, and future work seeks a partition that jointly uses spatial information and land-use homogeneity.The authors also propose adding periodic demand and points of interest as further information.
Loading 1905.06335v1…