Source-linked AI summary
Deep Multi-View Spatial-Temporal Network for Taxi Demand Prediction
Huaxiu Yao, Fei Wu, Jintao Ke, Xianfeng Tang, Yitian Jia, Siyu Lu, Pinghua Gong, Jieping Ye, Zhenhui Li
TL;DR
Taxi demand prediction requires modeling complex spatial and temporal relations that traditional time-series methods and separate spatial or temporal approaches do not jointly capture. DMVST-Net combines temporal LSTM, local spatial CNN, and semantic graph views, and outperforms competing methods on large-scale taxi-demand data. The authors identify interpretability and explicit semantic information as areas for future work.
Problem
Taxi demand prediction must use large-scale request data while addressing spatial and temporal relations that existing methods often model separately.
Method
DMVST-Net jointly models temporal, local spatial, and semantic relations using LSTM, local CNN, and graph embedding.
Results
DMVST-Net significantly outperforms several competing methods on a large-scale taxi-demand dataset.
Takeaways & Limitations
The experiments support a unified multi-view approach for taxi demand prediction that combines spatial, temporal, and semantic information.
Takeaways & Limitations
The method’s semantic information is implicit, and the authors note that deep-learning improvements are difficult to interpret.
Abstract
from arXiv · showhide
Taxi demand prediction is an important building block to enabling intelligent transportation systems in a smart city. An accurate prediction model can help the city pre-allocate resources to meet travel demand and to reduce empty taxis on streets which waste energy and worsen the traffic congestion. With the increasing popularity of taxi requesting services such as Uber and Didi Chuxing (in China), we are able to collect large-scale taxi demand data continuously. How to utilize such big data to improve the demand prediction is an interesting and critical real-world problem. Traditional demand prediction methods mostly rely on time series forecasting techniques, which fail to model the complex non-linear spatial and temporal relations. Recent advances in deep learning have shown superior performance on traditionally challenging tasks such as image classification by learning the complex features and correlations from large-scale data. This breakthrough has inspired researchers to explore deep learning techniques on traffic prediction problems. However, existing methods on traffic prediction have only considered spatial relation (e.g., using CNN) or temporal relation (e.g., using LSTM) independently. We propose a Deep Multi-View Spatial-Temporal Network (DMVST-Net) framework to model both spatial and temporal relations. Specifically, our proposed model consists of three views: temporal view (modeling correlations between future demand values with near time points via LSTM), spatial view (modeling local spatial correlation via local CNN), and semantic view (modeling correlations among regions sharing similar temporal patterns). Experiments on large-scale real taxi demand data demonstrate effectiveness of our approach over state-of-the-art methods.
Introduction
The paper addresses taxi demand prediction for smart-city transportation by combining spatial, temporal, and semantic relations in a unified deep model. Experiments on large-scale Didi Chuxing data show consistent improvement over competing baselines.
- Accurate taxi demand prediction can support resource pre-allocation and reduce unnecessary energy consumption in urban transportation.
- Existing traffic-prediction approaches often model spatial or temporal relations independently, motivating a joint treatment of both.
- Local CNN excludes weakly correlated remote regions because including them can hurt prediction performance.
- Semantic graph modeling captures similarities between spatially distant regions with similar demand patterns.
- DMVST-Net jointly models spatial, temporal, and semantic relations through local CNN, LSTM, and region-graph representations.
- The method consistently outperforms competing baselines on a large-scale Didi Chuxing taxi-request dataset.
Related Work
Prior traffic-prediction research includes time-series, context-based, spatial, and sequential models, but the paper emphasizes their incomplete treatment of spatial and temporal interactions. DMVST-Net instead models both relations jointly.
- Traditional traffic prediction commonly uses time-series methods such as ARIMA and its variants.
- Some studies incorporate external context or spatial interactions, including venue types, weather, events, road connectivity, and spatial smoothing.
- Context-based deep models may use extensive features without modeling spatial and temporal interactions.
- CNN-based approaches capture spatial correlation by treating city-wide traffic as images, but may include irrelevant remote regions.
- LSTM-based approaches model sequential dependency but may omit spatial relations.
- The paper’s main distinction is a joint deep-learning model that considers spatial and temporal sequential relations together.
Preliminaries
The paper formulates taxi demand over partitioned city regions and fixed time intervals, then predicts the next interval from historical demand and optional context features. A prediction function is defined to capture spatial-temporal interactions.
- The city is partitioned into non-overlapping locations, and time is divided into intervals of 30 minutes.
- A taxi request is represented by its timestamp, location, and user identification number.
- Demand is the number of taxi requests recorded at one location during a time interval.
- The prediction task estimates demand at time interval t + 1 using data available through time interval t.
- Context features for each location and time point may supplement historical demand data.
- The prediction function operates over all regions and previous time intervals to capture complex spatial and temporal interactions.
Proposed DMVST-Net Framework
DMVST-Net combines spatial, temporal, and semantic views to predict taxi demand. Its local CNN focuses on nearby regions, LSTM models sequential dependencies with context, and graph embeddings capture similarities among functionally related regions.
- DMVST-Net uses three views—spatial, temporal, and semantic—to model complementary relationships in taxi demand.The views are jointly integrated for prediction.
- Spatial View: Local CNN: The local CNN represents each region with an S × S neighborhood image and applies shared convolutional layers followed by dimensionality reduction.Zero padding handles locations at city boundaries, and convolutional parameters are shared across regions.
- Temporal View: LSTM: The temporal view uses an LSTM to model sequential demand dependencies from spatial representations and context features at corresponding times.The LSTM maintains a memory cell and uses input, forget, and output gates to process sequential information.
- Semantic View: Structural Embedding: The semantic view constructs a fully connected graph whose edge weights represent demand-pattern similarity between regions.Similarity is measured with Dynamic Time Warping on average weekly demand series, with α = 1.
- Semantic View: Structural Embedding: Graph embedding converts each region into a low-dimensional vector, which is passed through a fully connected layer for joint training.The paper uses LINE to generate the embeddings.
- Prediction Component: The model concatenates the three views and feeds the resulting representation into a fully connected network to predict demand at t + 1.Predictions are normalized to [0, 1] and later denormalized to actual demand values.
- Loss Function: Training jointly minimizes mean square loss and squared mean absolute percentage loss, using the same loss for all compared regression methods.The combined objective balances errors on large-valued samples with percentage-based error.
Experiment
Experiments on large-scale Guangzhou taxi-request data evaluate DMVST-Net against established baselines and model variants. Results show that combining temporal, local spatial, and semantic views improves prediction accuracy and robustness across days and neighborhood sizes.
- Dataset and setup: The Guangzhou dataset covers 20 × 20 regions, with 0.7km × 0.7km cells and about 300,000 requests per day.Data span 02/01/2017–03/26/2017; 47 days were used for training and 7 for testing, with 30-minute intervals.
- Baseline comparison: DMVST-Net achieves the lowest MAPE (0.1616) and RMSE (9.642), improving over the best baselines by 12.17% and 3.70%, respectively.The compared regression methods use context features and the same loss function, but do not model temporal and spatial dependency.
- Baseline comparison: DMVST-Net improves over ST-ResNet by 18.01% in MAPE and 6.37% in relative performance.The paper attributes this comparison to additionally modeling temporal dependency, context features, local spatial correlations, and semantic region similarity.
- Ablation study: Adding neighboring spatial features reduces MAPE by 0.63%, while the local CNN variant reduces it by 6.10%; combining all views performs best.The semantic view also reaches MAPE 0.1708 and RMSE 9.789, improving over the temporal-only variant.
- Robustness analysis: DMVST-Net consistently outperforms the displayed competing methods across all seven days, while weekend predictions are generally worse than weekday predictions.Its weekend-to-weekday relative error increase is smallest at 4.04%; the paper relates weekend difficulty to less regular demand patterns.
- Parameter analysis: A 4-hour input sequence gives the best performance, while larger temporal windows remain mostly stable but slightly degrade.For the local CNN, a 9 × 9 neighborhood with three convolutional layers performs best; enlarging the map to 13 × 13 increases prediction error by more than 3.46%.
Conclusion and Discussion
DMVST-Net integrates spatial, temporal, and semantic views through local CNN, LSTM, and semantic graph embedding for taxi demand prediction. Experiments report significant improvements over competing methods, while interpretability remains an open issue.
- DMVST-Net integrates spatial, temporal, and semantic views through local CNN, LSTM, and semantic graph embedding.
- The proposed method significantly outperforms several competing methods.
- The authors identify deep-learning interpretability as an unresolved issue and plan to investigate improvement sources and incorporate explicit information such as POI data.