Artificial I Intelligence f for S Smart Transp sportation Yan - - PowerPoint PPT Presentation

▶
artificial i intelligence f for s smart transp sportation
SMART_READER_LITE
LIVE PREVIEW

Artificial I Intelligence f for S Smart Transp sportation Yan - - PowerPoint PPT Presentation

Jun 17, 2023 161 likes •715 views

Artificial I Intelligence f for S Smart Transp sportation Yan Liu Associate Professor Computer Science Department University of Southern California Yan Liu (USC) ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE

slide-1
SLIDE 1

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Artificial I Intelligence f for S Smart Transp sportation

Yan Liu

Associate Professor Computer Science Department University of Southern California

ICML Time Series Workshop Yan Liu (USC) 1 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-2
SLIDE 2

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

AI and Machine Learning

Neural Networks Machine Learning: supervised, unsupervised Deep Learning Reinforcement Learning

?

ICML Time Series Workshop Yan Liu (USC) 2 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-3
SLIDE 3

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

GPS Data

Location Data and Floating-Car Trajectory

ICML Time Series Workshop Yan Liu (USC) 3 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-4
SLIDE 4

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Sensors

Loop detector, camera, microphone, mobile sensors …

ICML Time Series Workshop Yan Liu (USC) 4 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-5
SLIDE 5

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Transportation AI

Big data makes AI possible for transportation.

ICML Time Series Workshop Yan Liu (USC) 5 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-6
SLIDE 6

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Smart Transportation Brain

Control Analysis Data Collection

Signal Control Freeway Control Traffic Guidance Incident Management AI Dispatch Performance Measures Congestion Diagnosis Network Design Traffic Simulation Accident Analysis Ride-sharing Company Data Government Data Collaborators’ Data Crowd Sourced Data

S u p p l y D e m a n d

  • Ride-sharing

Services Platform Optimization Map Services

Taxi Express Car Pool Premiere …… Demand-Supply Prediction Order Dispatch Car Pooling Resource Allocation Multi-modal Route Planning ETA Pick-up locations VR Navigation Route Planning

ICML Time Series Workshop Yan Liu (USC) 6 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-7
SLIDE 7

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Outline

  • Traffic estimation and forecasting
  • Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic

Forecasting, ICLR 2018

  • Demand forecasting
  • Li et al, Spatiotemporal Multi-Graph Convolution for Ride-hailing Demand

Forecasting, AAAI 2019

  • Multi-rate multi-resolution forecasting/interpolation
  • Che et al, Hierarchical Deep Generative Models for Multi-Rate Multivariate Time

Series, ICML 2018

ICML Time Series Workshop Yan Liu (USC) 7 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-8
SLIDE 8

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Traffic Prediction

  • Input: road network and past T’ traffic speed observed at sensors
  • Output: traffic speed for the next T steps

7:00 AM 8:00 AM Input: Observations Output: Predictions

... ...

8:10AM, 8:20AM, …, 9:00 AM

ICML Time Series Workshop Yan Liu (USC) 8 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-9
SLIDE 9

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Existing Work

  • KNN-based models
  • Time series models
  • Seasonal Autoregressive Integrated Moving Average (S-ARIMA)
  • Support vector regression
  • Our prior work:
  • Latent space models: Dingxiong Deng et al, Latent Space Model for Road Networks

to Predict Time-Varying Traffic. KDD, 2016

  • Mixture LSTM: Y. Qi et al, Deep Learning: A Generic Approach for Extreme

Condition Traffic Forecasting. SDM 2016

ICML Time Series Workshop Yan Liu (USC) 9 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-10
SLIDE 10

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Challenges for Traffic Forecasting

Complex Spatial Dependency

Speed (mile/h)

Non-linear, non-stationary Temporal Dynamic

ICML Time Series Workshop Yan Liu (USC) 10 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-11
SLIDE 11

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Challenges for Traffic Forecasting

  • Spatial relationship among traffic flow is non-Euclidean and directed

ICML Time Series Workshop Yan Liu (USC) 11 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-12
SLIDE 12

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Traffic Forecasting with Convolution on Graph

  • Model spatial dependency with proposed diffusion convolution on graph

* Yagu guang g Li et al, Diffusion Convolutional Recurrent Neural Network: Data-dr driven n Traffic Forecasting

  • ng. ICLR, 2018

ICML Time Series Workshop Yan Liu (USC) 12 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-13
SLIDE 13

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Spatial Dependency in Traffic Prediction

  • Spatial dependency among traffic flow

Sensor 1 Sensor 2 Sensor 3

Close in Euclidean space Similar traffic speed

is no non-Eucl clidean and direct cted

!"#$%&' () → (+ ≠ !"#$%&' () → (+

ICML Time Series Workshop Yan Liu (USC) 13 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-14
SLIDE 14

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Spatial Dependency Modeling

  • Model the network of traffic sensors, i.e., loop detectors, as a

directed graph

  • Graph ! = ($, &)
  • Vertices (: o sensors
  • Adjacency matrix &: → weight between vertices

*+, = exp − dist567 8+, 8,

9

:9 if dist567 8+, 8, ≤ = dist567 8+, 8, : road network distance from 8+ to 8,, =: threshold to ensure sparsity, :9 variance of all pairwise road network distances

ICML Time Series Workshop Yan Liu (USC) 14 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-15
SLIDE 15

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Problem Statement

  • Graph signal: !" ∈ ℝ|&|×(, observation on ) at time *
  • + : number of vertices
  • , : feature dimension of each vertex.
  • Problem Statement: Learn a function -(·) to map 12 historical graph

signals to future 1 graph signals

… …

!456789 !4 !489 !486

  • .

ICML Time Series Workshop Yan Liu (USC) 15 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-16
SLIDE 16

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Generalize Convolution to Graph

  • Diffusion convolution filter: combination of diffusion processes with different

steps on the graph.

Max Min Filter weight

= !" + !# + !$ + … + !%

0 Step Diffusion 1 Step Diffusion 2 Step Diffusion K Step Diffusion Example diffusion filter Centered at &:,) ⋆+ ,

  • = /

01" %2#

!0 34

256 0 &:,)

Transition matrices of the diffusion process Learning complexity: 7 8

⋆+ : diffusion convolution, 9:: diagonal out-degree matrix.

ICML Time Series Workshop Yan Liu (USC) 16 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-17
SLIDE 17

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Generalize Convolution to Graph

  • Diffusion convolution filter: combination of diffusion processes with

different steps on the graph.

= !" + !# + !$ + … + !%

0 Step Diffusion 1 Step Diffusion 2 Step Diffusion K Step Diffusion Example diffusion filter Centered at &:,) ⋆+ ,

  • = /

01" %2#

!0,# 34

256 0 + !0,$ 38 256⊺ 0 &:,)

Dual directional diffusion to model upstream and downstream separately

⋆+ : diffusion convolution, :;: diagonal out-degree matrix, :<: diagonal in-degree matrix Max Min weight

ICML Time Series Workshop Yan Liu (USC) 17 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-18
SLIDE 18

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Advantage of Diffusion Convolution

  • Efficient
  • Learning complexity: ! "
  • Time complexity: ! " # , # number of edges
  • Expressive
  • Many popular convolution operations, including the ChebNet [Defferrard et al.,

NIPS ’16], can be seen as special cases of the diffusion convolution [Li et al. ICLR ’18]. %:,' ⋆) *

+ = - ./0 123

4.,3 56

278 . + 4.,: 5; 278⊺ . %:,'

* Defferrard, M et al, Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering, NIPS, 2016 * Yaguang Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic Forecasting, ICLR, 2018 ⋆) : diffusion convolution, =>: diagonal out-degree matrix, =?: diagonal in-degree matrix

ICML Time Series Workshop Yan Liu (USC) 18 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-19
SLIDE 19

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Diffusion Convolutional Recurrent Neural Network

  • Diffusion Convolutional Recurrent Neural Network (DCRNN)
  • Model spatial dependency with diffusion convolution
  • Sequence to sequence learning with encoder-decoder framework

* Yaguang Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic Forecasting, ICLR 2018

ICML Time Series Workshop Yan Liu (USC) 19 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-20
SLIDE 20

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Multi-step ahead prediction with RNN

! x#

DCGRU

! x$

DCGRU

! x%

DCGRU

Previous model

  • utput is fed into

the network

! x# ! x$

Error Propagation

! x x

Model prediction Observation or ground truth

DCGRU

x&

DCGRU

x'

DCGRU

x(

Teach the model to deal with its own error.

Current Time

Model Temporal Dynamics using Recurrent Neural Network

ICML Time Series Workshop Yan Liu (USC) 20 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-21
SLIDE 21

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Improve Multi-step ahead Forecasting

  • Traffic prediction as a sequence to sequence learning problem
  • Encoder-decoder framework

DC DCGRU

!"

DC DCGRU

x$

DC DCGRU

!%

Enc Encoder der

& x'

DC DCGRU

& x(

DC DCGRU

& x)

DC DCGRU

De Decoder

<G <GO> O>

!' !(

& ! !

Model prediction Observation or ground truth Current Time

!' !( !)

Backprop errors from multiple steps.

* Sutskever et al. Sequence to sequence learning with neural networks, NIPS 2014 Ground truth becomes unavailable in testing.

*' *( *)

!", !$, !% → !' !", !$, !% → !', !(, !)

ICML Time Series Workshop Yan Liu (USC) 21 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-22
SLIDE 22

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Improve Multi-step ahead Forecasting

  • Improve multi-step ahead forecasting with scheduled sampling

DC DCGRU x" DC DCGRU x# DC DCGRU x$ % x& DC DCGRU % x' DC DCGRU % x( DC DCGRU <G <GO> O> x& % x& x' % x'

Sc Scheduled sampling: g: Choose to use the previous gr ground t und trut uth h or mo model l prediction by flipping a coin

% ) )

Model prediction Observation or ground truth

Enc Encoder der De Decoder

Current Time * Bengio,Samy et al. Scheduled sampling for sequence prediction with recurrent neural networks. NIPS 2015

ICML Time Series Workshop Yan Liu (USC) 22 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-23
SLIDE 23

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Diffusion Convolutional Recurrent Neural Network

  • Diffusion Convolutional Recurrent Neural Network (DCRNN)
  • Model spatial dependency with diffusion convolution
  • Sequence to sequence learning with encoder-decoder framework
  • Improve multi-step ahead forecasting with scheduled sampling

* Yaguang Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic Forecasting, ICLR 2018

ICML Time Series Workshop Yan Liu (USC) 23 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-24
SLIDE 24

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Experiment - Datasets

  • METR-LA:
  • 207 traffic sensors in Los Angeles
  • 4 months in 2012
  • 6.5M observations
  • PEMS-BAY:
  • 345 traffic sensors in Bay Area
  • 6 months in 2017
  • 17M observations

ICML Time Series Workshop Yan Liu (USC) 24 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-25
SLIDE 25

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Experiments

  • Baselines
  • Historical Average (HA)
  • Autoregressive Integrated Moving Average (ARIMA)
  • Support Vector Regression (SVR)
  • Vector Auto-Regression (VAR)
  • Feed forward Neural network (FNN)
  • Fully connected LSTM with Sequence to Sequence

framework (FC-LSTM)

  • Task
  • Multi-step ahead traffic speed forecasting

ICML Time Series Workshop Yan Liu (USC) 25 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-26
SLIDE 26

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Experimental Results

  • DCRNN achieves the best performance for all forecasting

horizons for both datasets

1.00 2.00 3.00 4.00 5.00 6.00 7.00

15 Min 30 Min 1 Hour Mean Absolute Error (MAE)

METR-LA

HA ARIMA VAR SVR FNN FC-LSTM DCRNN

1.00 1.50 2.00 2.50 3.00 3.50 15 Min 30 Min 1 Hour Mean Absolute Error (MAE)

PEMS-BAY

HA ARIMA VAR SVR FNN FC-LSTM DCRNN

ICML Time Series Workshop Yan Liu (USC) 26 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-27
SLIDE 27

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Effects of Spatiotemporal Dependency Modeling

  • w/o temporal: removing sequence to sequence learning.
  • w/o spatial: remove the diffusion convolution.

1.5 2 2.5 3 3.5 4 4.5 5 15 Min 30 Min 1 Hour Mean Absolute Error (MAE)

DCRNN w/o Temporal DCRNN w/o Spatial DCRNN

Removing either spatial or temporal modeling results in significantly worse results.

METR-LA

ICML Time Series Workshop Yan Liu (USC) 27 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-28
SLIDE 28

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Outline

  • Traffic estimation and forecasting
  • Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic

Forecasting, ICLR 2018

  • Demand forecasting
  • Li et al, Spatiotemporal Multi-Graph Convolution for Ride-hailing Demand

Forecasting, AAAI 2019

  • Multi-rate multi-resolution forecasting/interpolation
  • Che et al, Hierarchical Deep Generative Models for Multi-Rate Multivariate Time

Series, ICML 2018

ICML Time Series Workshop Yan Liu (USC) 28 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-29
SLIDE 29

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 29 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Introduction More than 18 billion ride-hailing trips worldwide in 2018*

– Twice as much as the world population.

Benefit of better ride-hailing demand forecasting

Better Vehicle Dispatching Early congestion warning Higher vehicle utilization

* http://www.businessofapps.com/data/uber-statistics/, Nov 2018.

ICML Time Series Workshop Yan Liu (USC) 29 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-30
SLIDE 30

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 30 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Region-level Ride-hailing Demand Forecasting Input: past T observations of demands of all |"| regions Output: demands of all |"| regions in the next time stamp

Input Output

...

#: ℝ&×|(| → ℝ|(| ℝ&×|(| ℝ|(|

Complicated spatial and temporal correlations

ICML Time Series Workshop Yan Liu (USC) 30 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-31
SLIDE 31

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 31 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Related Work

Spatiotemporal forecasting on grid

– Classical settings for demand forecasting problem – CNN-based approaches: region-wise relationship is Euclidean

  • DeepST/STResNet: Crowd flow forecasting (Zhang et al., 2017)
  • DMVST: Demand forecasting (Yao et al., 2018)

Spatiotemporal forecasting on graph

– LinUOTD: handcrafted feature + LR for demand forecasting (Tong et al., 2017) – DCRNN/ST-GCN: Graph convolution based traffic forecasting (Li et al., 2018a, Yu et al., 2018, Li et al., 2018b, Yan et al., 2018)

Hard to capture the non-Euclidean correlations Hard to capture the multimodal correlations

ICML Time Series Workshop Yan Liu (USC) 31 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-32
SLIDE 32

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 32 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Multimodal Correlations among Regions Spatial proximity

– Region 1 and 2

Functional similarity

– Regions with similar context show similar demand patterns – Region 1 and 3

Road connectivity

– High-speed transportation facilitate correlation – Region 1 and 4

ICML Time Series Workshop Yan Liu (USC) 32 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-33
SLIDE 33

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Spatiotemporal Multi-Graph Convolution Network

Page 33

Re Reweight and aggregate temporal observations

RNN

Contextual Gated RNN

Contextual Gating

Encode pair-wise correlations using gr graphs

Spatial proximity

  • Func. similarity

Connectivity Region Correlation

Capture spatial correlations among regions with mu multi- gr graph convolution

GCN GCN GCN

… …

Graph convolution Graph convolution Graph convolution

…

Generate prediction Aggregation

ICML Time Series Workshop Yan Liu (USC) 33 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-34
SLIDE 34

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

CGRNN: Context-aware Temporal Aggregation

Page 34

  • !
  • [# $ , #($'(), … ] ∈ ℝ.× 0 ×1

! 2 ∈ ℝ.×3×1 4

5667

  • Summarize contextual

information

  • Calculate gates based on

interdependencies between observations with self-attention

  • Reweight observations

with gates

  • Aggregate reweighted
  • bservations with share-

weight RNN

  • 8

# $ , 8 # $'( , … ∈ ℝ.× 0 ×1 !

  • |:|
  • ℝ 0 ×1;
  • < ∈ ℝ.
  • ICML Time Series Workshop

Yan Liu (USC) 34 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-35
SLIDE 35

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Spatiotemporal Multi-Graph Convolution Network

Page 35

Generate prediction Encode pair-wise correlations using gr graphs

Spatial proximity

  • Func. similarity

Connectivity

Re Reweight and aggregate

  • bservations

RNN

Contextual Gated RNN

Contextual Gating

Capture spatial correlations among regions with mu multi- gr graph convolution

GCN GCN GCN

… …

Graph convolution Graph convolution Graph convolution

…

Region Correlation

Aggregation

ICML Time Series Workshop Yan Liu (USC) 35 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-36
SLIDE 36

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 36 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Multi-graph Convolution

! = MGC & = ' Agg * +,; ., &/

A, ∈ 1

& ∈ ℝ 3 ×5

! ∈ ℝ 3 ×56

= '

/ ∈ ℝ5×56 Feature transformation * +,; ., ∈ ℝ 3 ×|3| Node aggregation

Agg

A, ∈ 1 Aggregation function

* +,; ., : function of adjacency matrix +, with parameter .,

– Polynomial of graph Laplacian, graph attention etc.

Agg: Aggregation function

– Sum, average, attention-based aggregation

ICML Time Series Workshop Yan Liu (USC) 36 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-37
SLIDE 37

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 37 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Datasets Beijing:

– 1296 regions, 19M samples – 10 months in 2017

Shanghai

– 896 regions, 13M samples – 10 months in 2017

POI/Road network

– OpenStreetMap

Beijing Shanghai

ICML Time Series Workshop Yan Liu (USC) 37 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-38
SLIDE 38

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 38 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Experiments Baselines

– Historical Average (HA) – Linear Regression (LASSO, Ridge) – Vector Auto-Regression (VAR) – Spatiotemporal Auto-Regressive Model (STAR) – Gradient Boosted Machine (GBM) – Spatiotemporal Residual Network (ST-ResNet), with Euclidean grid – Spatiotemporal graph convolutional network (ST-GCN), with road network graph – Deep Multi-view Spatiotemporal Network (DMVST-Net), with Euclidean grid, SO SOTA for ride-ha hailing ng dem demand nd fo forecasting

Task

– One step ahead ride-hailing demand forecasting

ICML Time Series Workshop Yan Liu (USC) 38 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-39
SLIDE 39

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 39 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

6.0 8.0 10.0 12.0 14.0 16.0 18.0

Root Mean Square Error

HA LASSO Ridge VAR STAR GBM STResNet DMVST-Net ST-GCN ST-MGCN

Experimental Results ST-MGCN achieves the be best t pe performanc nce on both datasets

– 10+% improvement*.

Beijing Shanghai

* In terms of relative error reduction of RMSE.

` `

ICML Time Series Workshop Yan Liu (USC) 39 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

10+% improvement on real-world large-scale datasets

slide-40
SLIDE 40

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ICML Time Series Workshop 40 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Experimental Results Both spatial and temporal correlations modeling are necessary

– Removing either graph component leads to si signifi ficantly worse se performance. – With CG CGRNN, ST-MGCN achieves the best performance.

10.0 10.5 11.0 11.5 12.0 12.5 13.0

Root Mean Square Error

w/o Spatial Proximity w/o Functional w/o Transportation ST-MGCN

10.0 10.5 11.0 11.5 12.0 12.5 13.0

Root Mean Square Error

Average Pooling CG RNN CG + RNN

Effect of spatial correlation modeling Effect of temporal correlation modeling

ICML Time Series Workshop Yan Liu (USC) 40 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-41
SLIDE 41

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Outline

  • Traffic estimation and forecasting
  • Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic

Forecasting, ICLR 2018

  • Demand forecasting
  • Li et al, Spatiotemporal Multi-Graph Convolution for Ride-hailing Demand

Forecasting, AAAI 2019

  • Multi-rate multi-resolution forecasting/interpolation
  • Che et al, Hierarchical Deep Generative Models for Multi-Rate Multivariate Time

Series, ICML 2018

ICML Time Series Workshop Yan Liu (USC) 41 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-42
SLIDE 42

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Outline

  • Traffic estimation and forecasting
  • Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic

Forecasting, ICLR 2018

  • Demand forecasting
  • Li et al, Spatiotemporal Multi-Graph Convolution for Ride-hailing Demand

Forecasting, AAAI 2019

  • Multi-rate multi-resolution forecasting/interpolation
  • Che et al, Hierarchical Deep Generative Models for Multi-Rate Multivariate Time

Series, ICML 2018

ICML Time Series Workshop Yan Liu (USC) 42 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-43
SLIDE 43

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Introduction

  • Multivariate Time Series (MTS) -- many real-world applications
  • Healthcare
  • One of the key challenges
  • Different sampling rates
  • Multiple data sources / sensors

, climate, traffic, financial forecasting, engineering…

  • - Multi-Rate Multivariate Time Series (MR-MTS)

ICML Time Series Workshop Yan Liu (USC) 43 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-44
SLIDE 44

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Motivation

  • Major challenges of modeling MR-MTS
  • Need to handle different sampling rates
  • Multi-scale temporal dependencies
  • Complex underlying generation mechanism
  • Existing solutions to MR-MTS forecasting/interpolation problems
  • Single-rate model?

(Kalman filter, VAR, deep Markov models, ...)

  • Ignoring dependencies across different rates
  • Simple imputations?

(mean-imputation, Spline, MICE, MissForest, ...)

  • May introduce unrelated/hide necessary dependencies
  • Multi-rate discriminative models?

(PLSTM, HM-RNN, ...)

  • Not able to learn how the data is generated

ICML Time Series Workshop Yan Liu (USC) 44 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-45
SLIDE 45

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Motivation

  • Major challenges of modeling MR-MTS
  • Need to handle different sampling rates
  • Multi-scale temporal dependencies
  • Complex underlying generation mechanism
  • Key point
  • To learn the latent hierarchical structures of the data generation mechanism
  • Our proposed solution
  • MR-HDMM: Multi-Rate Hierarchical Deep Markov Model

ICML Time Series Workshop Yan Liu (USC) 45 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-46
SLIDE 46

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Overview

  • Problem definitions
  • Input -- MR-MTS of ! different sampling rates and " time steps (#$:&

$:')

  • Case 1 -- Forecasting problem
  • Output -- Given ():*

):+, predict (*:*, ):+

  • Case 2 -- Interpolation problem
  • Output -- Fill-in missing values of lower sampling rates in ():*

):+

  • MR-HDMM: Multi-Rate Hierarchical Deep Markov Model
  • Component -- a generation model and an inference model
  • Motivation -- capturing hierarchical structures in underlying data generation process
  • Learnable switches
  • Auxiliary connections

ICML Time Series Workshop Yan Liu (USC) 46 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-47
SLIDE 47

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Generation Model

  • Transition
  • Learning latent states !
  • To capture hierarchical structure
  • Learnable switches
  • Update-and-reuse
  • Emission
  • Generating MR-MTS "
  • To capture multi-scale dependencies
  • Auxiliary connections

Solving marginal MLE?

ICML Time Series Workshop Yan Liu (USC) 47 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-48
SLIDE 48

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Inference and Learning

  • Keep similar structure as the generative model
  • Keeping the Markov properties of !
  • Inheriting the same switches "
  • Capturing MR-MTS observation by multiple RNNs
  • Maximize the variational evidence lower bound

(ELBO)

  • Conditional likelihood
  • KL at each time step and for each layer

Jointly learning all parameters by stochastic backpropagation and ancestral sampling

#

$%& '

#

(%& )

*+∗(./

0:2) 456 789 :$

( |.$ &:(

#

$%& '

*+∗(./<0

)=>) ?@ .$ &|:&:' &:), .$B& &

||78 .$

&|.$B& &

+ #

$%& '

#

(%D )

*+∗(./<0

2

,./

2<0)=>) ?@ .$

(|:&:' &:), .$B& (

, .$

(B& ||78 .$ (|.$B& (

, .$

(B&

ICML Time Series Workshop Yan Liu (USC) 48 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-49
SLIDE 49

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Experimental settings

  • Datasets
  • MIMIC-III: 5 runs × 5-fold CV (train/valid/test split)
  • USHCN: 5 runs of train/valid/test split with 1-month stride
  • Forecasting baselines
  • Single-rate: Kalman Filter, VAR, Deep Markov Model, HM-RNN, LSTM, and PLSTM
  • Multi-rate: Multiple KF, Multi-Rate KF, and two simplified models of MR-HDMM
  • Interpolation baselines
  • Imputation: Mean, CubicSpline, MICE, MissForest, SoftImpute
  • Deep learning: Deep Markov Model and the two simplified models of MR-HDMM

Domain Dataset # of Samples Sampling Rates # of Variables Time Series Length Healthcare MIMIC-III 10709 (admissions) 1 / 4 / 12 Hours 7 / 12 / 44 72 Hours Climate USHCN 100 (years) 1 / 5 / 10 Days 70 / 69 / 69 365 Days

ICML Time Series Workshop Yan Liu (USC) 49 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-50
SLIDE 50

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Quantitative results

  • Forecasting
  • Interpolation

HSR/MSR/LSR: High/Mid/Low sampling rate In/Out-Sample: Interpolating training/testing dataset

ICML Time Series Workshop Yan Liu (USC) 50 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-51
SLIDE 51

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Visualizations of the learned latent hierarchical structures

  • First 48 hours of an admission from MIMIC-III dataset
  • Blue: update of higher-layer states (!")
  • Red: update of lower-layer states (!#)
  • Higher layer ⇒ fewer updates ⇒ longer-term dependencies
  • A 1-year climate observation from USHCN dataset
  • Green: precipitation records
  • Changes in precipitations ⇒ significant differences ⇒ captured by the higher layer

ICML Time Series Workshop Yan Liu (USC) 51 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-52
SLIDE 52

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Outline

  • Traffic estimation and forecasting
  • Li et al. Diffusion Convolutional Recurrent Neural Network: Data-driven Traffic

Forecasting, ICLR 2018

  • Demand forecasting
  • Li et al, Spatiotemporal Multi-Graph Convolution for Ride-hailing Demand

Forecasting, AAAI 2019

  • Multi-rate multi-resolution forecasting/interpolation
  • Che et al, Hierarchical Deep Generative Models for Multi-Rate Multivariate Time

Series, ICML 2018

ICML Time Series Workshop Yan Liu (USC) 52 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-53
SLIDE 53

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Open Dataset

KDD Cup 2017 Highway Tollgates Traffic Flow Prediction GAIA Open Dataset Trajectory and OD data Uber Movement Federal Highway Administration Next Generation Simulation (NGSIM) Program Public Data

ICML Time Series Workshop Yan Liu (USC) 53 ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

slide-54
SLIDE 54

ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION ARTIFICIAL INTELLIGENCE FOR SMART TRANSPORTATION

Acknowledgement

  • PhD students and Postdoc: Yaguang Li (Google), Zhengping Che (DiDi),

Rose Yu (Northeastern), Sanjay Purushotham (U of Maryland, Baltimore)

  • Collaborators
  • Funding agency

ICML Time Series Workshop Yan Liu (USC) 54