1. Passenger Demand Prediction with Cellular Footprints
Jing Chu1, Kun Qian1, Xu Wang1, Lina Yao2, Fu Xiao3, Jianbo Li4,
Xin Miao1 and Zheng Yang1
1School of Software, Tsinghua University
2School of Computer Science and Engineering, The University of New South Wales
3School of Computer Science, Nanjing University of Posts and Telecommunications
4College of Computer Science and Techonology, Qingdao University
Presenter: Jing Chu

2. Motivation Imbalance Passenger Demand Prediction
Passenger Waiting Time
Driver Profit

3. Motivation Cellular Data Can get crowd related data compared with
User Location
Mobility Modelling
Cellular Data
Data Traffic Engineering
User Portrait Research
Can get crowd related data compared with
the data from online car-hailing platform

4. Related Work Data sparsity problem Lack of crowd flow analysis
Target to the number of pick-ups, ignoring potential passengers who eventually give up taking taxis
Lack of crowd flow analysis
Reflect potential passenger demand
Improper handling of spatial relations
Fixed grid region
Fail to capture the complex spatial dependency
Using deep learning model to predict citywide passenger demand
with cellular data on flexible region partition accurately

5. Challenges Challenge 1 How to identify region characteristics?
How to realize the reasonable region partition?
Challenge 3
How to capture spatial and temporal correlation?

6. System Overview Database server module Pre-processing module
Deep learning module
Visualization and evaluation module

7. System Overview
Database server module: Stores the big cellular data and provides retrival and aggregation services for fast preprocessing
Pre-processing module
Deep learning module
Visualization and evaluation module

8. 290 Billion Cellular Records
Data Source
Cellular data: Collected by a major cellular carrier of China
Record: Unique anonymized user ID, create time, cell tower ID, App ID, URL
Dec 5th, 2016-Feb 4th 2017
290 Billion Cellular Records
8000 Cell Towers
1.5 Million Covered Users
Shenyang City, China
Weather data: From Dark Sky API
Weather States
Temperature
Wind Speed
Visibility

9. System Overview Database server module
Pre-processing module: Passenger information extraction and flexible region partition
Deep learning module
Visualization and evaluation module

10. Passenger Demand Extraction
Intercepting analysis of network data packets: Using HTTP proxy tool to understand the meaning behind URLs of DiDi Chuxing
Extract passenger demand from cellular data

11. Region Partition
Partition the city by primary road network from OpenStreetMap
Finer-grained partition: Partition the city by the main secondary
roads if the passenger demand of any block is too high

12. Crowd Outflow Extraction
Crowd outflow: The number of people leaving the region
Outflow
Start
Other regions
A region

13. System Overview Database server module Pre-processing module
Deep learning module: Deep learning architecture FlowFlexDP to
model and predict the passenger demand for each region
Visualization and evaluation module

14. Deep Learning Architecture FlowFlexDP
Predict the passenger demand for each region

15. Graph Convolutional Neural Network
Model spatial dependency: Graph Convolutional Neural Network
Spectral
Graph Theory
The normalized graph Laplacian
Convolution theorem
Polynomial filter: Consider k-order neighbors

16. GCNN with Residual Learning
Capture the long-distance spatial dependency:
Apply residual learning to GCNN
Residual learning
The residual unit of GCNN

17. Sequences Fusion Passenger demand = Hourly + Daily Closeness
Periodicity
Crowd Outflow Sequences Fusion
Cross Correlation Coefficient
Office Region
Residential Region

18. Parametric-matrix Based Fusion
External Factors Fusion
Predicted
Passenger
Demand
Weather Σ
Regression
Passenger Demand
Crowd Outflow
Weather & Time Metadata
Parametric-matrix Based Fusion
Time Metadata
Holiday
Snowy Day

19. System Overview Database server module Pre-processing module
Deep learning module
Visualization and evaluation module

20. Experimental Setting Training dataset
The first 65% for training data, the second 10% for validation set and the rest 25% for testing
Spatio-temporal sequence: Min-Max normalization to [0,1]
External factors: One-hot coding for day-of-week, time-of-day, holidays, weather state; Min-Max normalization for the temperature, wind speed and visibility
The length of hourly sequence: {3, 4, 5, 6, 7, 8}
The length of daily sequence: {1, 2, 3, 4, 5, 6, 7, 8}

21. Evaluation Baselines HA: The Historical Average model
ARIMA: The Autoregressive Integrated Moving Average
SARIMA: The Seasonal Autoregressive Integrated Moving Average model
VAR: Vector Auto-Regressive
LSTM: Long-Short Term Memory is a Recurrent Neural Network architecture
ANN: The Artificial Neural Network
Metric: RMSE

22. Results Performance Evaluation
ResGCNN: Combines GCNN and residual learning
ResGCNN-D: Further adds the passenger demand daily sequence
ResGCNN-DO: Further fuses the crowd outflow hourly and daily sequences
FlowFlexDP: Our final model
All variants achieve better performance than the baseline by at least 12.99%

23. Visualization
Our prediction results are very close to the real state

24. Conclusion We demonstrate cellular data as a rich data source for
passenger demand prediction, which has been largely
overlooked and unexplored previously
It is the first work that uses crowd flow information from
cellular data for passenger demand prediction
We propose a deep learning model FlowFlexDP, which uses
GCNN and residual learning for predicting passenger demand
on flexible region partition
Evaluation results on a large-scale real data set show that
our model outperforms existing models

25. Thanks
Q&A