Jing Chu 1 , Kun Qian 1 - PowerPoint Presentation

Slide1

Jing Chu1, Kun Qian1, Xu Wang1, Lina Yao2, Fu Xiao3, Jianbo Li4, Xin Miao1 and Zheng Yang11School of Software, Tsinghua University2School of Computer Science and Engineering, The University of New South Wales3School of Computer Science, Nanjing University of Posts and Telecommunications4College of Computer Science and Techonology, Qingdao UniversityPresenter: Jing Chu

Passenger Demand Prediction with Cellular Footprints

1

Slide2

Motivation2Imbalance Passenger Demand PredictionPassenger Waiting TimeDriver Profit

Slide3

MotivationCellular DataUser LocationMobility ModellingData Traffic EngineeringUser Portrait Research

Can get

crowd related data

compared with

the data from online car-hailing platform

Slide4

Related WorkData sparsity problemLack of crowd flow analysisTarget to the number of pick-ups, ignoring potential passengers who eventually give up taking taxisImproper handling of spatial relationsFixed grid regionFail to capture the complex spatial dependencyReflect potential passenger demandUsing deep learning model to predict citywide passenger demand with

cellular data on flexible region partition accurately

Slide5

Challenges

Slide6

System OverviewDatabase server module Pre-processing moduleDeep learning moduleVisualization and evaluation module

Slide7

System OverviewDatabase server module: Stores the big cellular data and provides retrival and aggregation services for fast preprocessingPre-processing moduleDeep learning moduleVisualization and evaluation module

Slide8

Data SourceCellular data: Collected by a major cellular carrier of China Weather data: From Dark Sky API290 Billion Cellular Records1.5 Million Covered UsersDec 5th, 2016-Feb 4th 20178000 Cell TowersShenyang City, ChinaWeather StatesTemperature

Wind SpeedVisibility

Record: Unique

anonymized user ID,

create

time,

cell

tower ID, App ID, URL

Slide9

System OverviewDatabase server modulePre-processing module: Passenger information extraction and flexible region partition Deep learning moduleVisualization and evaluation module

Slide10

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

Slide11

Region PartitionPartition the city by primary road network from OpenStreetMapFiner-grained partition: Partition the city by the main secondaryroads if the passenger demand of any block is too high

Slide12

Crowd Outflow ExtractionCrowd outflow: The number of people leaving the regionOutflowStartOther regionsA region

Slide13

System OverviewDatabase server modulePre-processing moduleDeep learning module: Deep learning architecture FlowFlexDP to model and predict the passenger demand for each regionVisualization and evaluation module

Slide14

Deep Learning Architecture FlowFlexDPPredict the passenger demand for each region

Slide15

Graph Convolutional Neural NetworkModel spatial dependency: Graph Convolutional Neural NetworkSpectral Graph Theory The normalized graph LaplacianConvolution theoremPolynomial filter: Consider k-order neighbors

Slide16

GCNN with Residual LearningCapture the long-distance spatial dependency: Apply residual learning to GCNNResidual learningThe residual unit of GCNN

Slide17

Sequences FusionPassenger demand = Hourly + DailyClosenessPeriodicityCrowd Outflow Sequences FusionCross Correlation CoefficientOffice RegionResidential Region

Slide18

External Factors FusionWeather & Time MetadataHolidaySnowy DayPredicted PassengerDemandWeatherΣRegressionPassenger Demand

Crowd Outflow

Time Metadata

Parametric-matrix Based Fusion

Slide19

System OverviewDatabase server modulePre-processing moduleDeep learning moduleVisualization and evaluation module

Slide20

Experimental SettingTraining datasetThe first 65% for training data, the second 10% for validation set and the rest 25% for testingSpatio-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 visibilityThe length of hourly sequence: {3, 4, 5, 6, 7, 8}The length of daily sequence: {1, 2, 3, 4, 5, 6, 7, 8}

Slide21

EvaluationBaselinesHA: The Historical Average modelARIMA: The Autoregressive Integrated Moving AverageSARIMA: The Seasonal Autoregressive Integrated Moving Average modelVAR: Vector Auto-RegressiveLSTM: Long-Short Term Memory is a Recurrent Neural Network architectureANN: The Artificial Neural NetworkMetric: RMSE

Slide22

ResultsPerformance EvaluationResGCNN: Combines GCNN and residual learningResGCNN-D: Further adds the passenger demand daily sequenceResGCNN-DO: Further fuses the crowd outflow hourly and daily sequencesFlowFlexDP: Our final modelAll variants achieve better performance than the baseline by at least 12.99%

Slide23

VisualizationOur prediction results are very close to the real state

Slide24

ConclusionWe propose a deep learning model FlowFlexDP, which usesGCNN and residual learning for predicting passenger demandon flexible region partitionEvaluation results on a large-scale real data set show that our model outperforms existing modelsWe demonstrate cellular data as a rich data source forpassenger demand prediction, which has been largelyoverlooked and unexplored previouslyIt is the first work that uses crowd flow information fromcellular data for passenger demand prediction

Slide25

25Thanks Q&A