1

Learning Transportation Mode from Raw GPS Data for Geographic Applications on the Web

Yu Zheng, Like Liu, Xing Xie, WWW’08Microsoft Research Asia

Advisor: Chia-Hui ChangPresenter: Teng-Kai Fan

Date: 2010-03-19

2

Outline

• Introduction• Framework• Methodology• Experiment• Conclusion & future work

3

2004 2005 2006 2007 2008 2009 2010 2011 -

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1.1

AfricaAsia/PacificEastern EuropeJapanLatin AmericaMiddle EastNorth AmericaWestern EuropeTotal

Years

Perc

enta

ge o

f Tot

al S

ales

• Background Percentage of GPS-enabled handset among mobile phone (Gartner Dataqueste: Forecast: GPS-enabled device 2004-2011)

4

Introduction• What we do: Infer transportation modes from users’ GPS logs

GPS log

Users

Infer model

5

Introduction– Motivation

• Differentiate GPS trajectory of different transportation modes• Learning knowledge from raw GPS data

– enable people to absorb more knowledge from others’ life experience– Trigger people’s memory about their past– Understand people’s life pattern

• Understanding user behavior– Context-aware computing– Modeling traffic condition– Discover social pattern– …

– Difficulty• A trajectory may contain more than two kinds of transportation modes• Pure velocity-based method may suffer from congestion

6

Introduction

0%

10%

20%

30%

40%

50%

60%

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

WalkCarBusBike

Distribution of mean velocity (m/s) of different transportation modes

0%

5%

10%

15%

20%

25%

30%

35%

40%

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Walk

Car

Bus

Bike

Distribution of maximum velocity (m/s) of different transportation modes

7

Introduction

• Contributions– We propose

• A change point-based segmentation method• An inference model based on supervised learning• A post-processing algorithm based on conditional probability

– Significance• A step toward mining knowledge from raw GPS data for geographic applications on

the Web• A step toward understanding user behavior based on GPS data

– Evaluation results• Large-scale data collected by 45 people over a period of 6 months• Almost 70 percent accuracy

8

Framework

Latitude, longitude, TimeP1: Lat1, long1, T1P2: Lat2, long2, T2 ………...Pn: Latn, longn, Tn

P1

Pn

Car

P2 P3 Pn-1

Change Point

WalkNon-Walk Segment

L2,T2

Walk Segment

• Preliminary

a place where people changetheir transportation modes

9

Framework

Inference ModelPost Process

GPS Log Segmentation

Feature Extraction

Features

Trans. Modes

Final Results

CRF

• Inference strategy

divide the GPS track into trips and then partition each trip into segments by change

points

10

Framework

Segment[i-1]: Car Segment[i]: Walk Segment[i+1]: Bike

P(Car): 75%P(Bus): 10%P(Bike): 8%P(Walk): 7%

P(Bike): 62%P(Walk): 24%P(Bus): 8%P(Car): 6%

P(Bike): 40%P(Walk): 30%P(Bus): 20%P(Car): 10%

Segment[i].P(Bike) = Segment[i].P(Bike) * P(Bike|Car)

Segment[i].P(Walk) = Segment[i].P(Walk) * P(Walk|Car)

• Post-Processing

11

Framework

Mi-1 Mi Mi+1

Xi-1 Xi Xi+1

Observations

States

WalkBus ForwardCar

Graphical Model

A Trip

• CRF-Based Inference

transportationsmode

feature from segment

12

Methodology• Commonsense knowledge from real world

– Typically, people need to walk before transferring transportation modes– Typically, people need to stop and then go when transferring modes– Walk should be a transition between different transportation modes

Transportation modes

Walk Car Bus Bike

Walk / 53.4% 32.8% 13.8%

Car 95.4% / 2.8% 1.8%

Bus 95.2% 3.2% / 1.6%

Bike 98.3% 1.7% 0% /

Transition matrix of transportation modes

13

Methodology• Change point-based Segmentation Algorithm

– Step 1: using a loose upper bound of velocity (Vt) and acceleration (at) to distinguish all possible Walk Points, non-Walk Points.

– Step 2: merge short segment (the length less than a thredshold) composed by consecutive Walk Points or non-Walk points

– Step 3: merge consecutive Uncertain Segment (less than 50 meters) to non-Walk Segment.

– Step 4: end point of each Walk Segment are potential change points

WalkBus

Certain Segment

Denotes a non-walk Point: P.V>Vt or P.a>at

Denotes a possible walk point: P.V<Vt and P.a<at

(b)

(c)

Backward Forward

Car

(a)

Certain Segment3 Uncertain Segments

Car

14

Experiments• Framework of experiment

• Feature Extraction– length– mean velocity– expectation of velocity– variance of velocity– top three velocities– top three accelerations

GPS log Data

Change Point Based

Uniform Duration Based

Uniform Length Based

Bayesian NetSVM Decision Tree CRF

Feature Extraction from Each Segment

Seg

men

tati

onIn

fere

nce

15

Experiment

• Devices

• Data

16

Experiment• Evaluation method

– Precision of inference a segment • Accuracy by Length

• Accuracy by Duration

– Change Point• Precision of change point• Recall of change point

𝐴𝐿 = σ 𝐶𝑜𝑟𝑟𝑒𝑐𝑡𝑆𝑒𝑔𝑚𝑒𝑛𝑡ሾ𝑗ሿ.𝐿𝑒𝑛𝑔𝑡ℎ𝑚𝑗=0σ 𝑆𝑒𝑔𝑚𝑒𝑛𝑡 ሾ𝑖ሿ.𝐿𝑒𝑛𝑔𝑡ℎ𝑁𝑖=0

𝐴𝐷 = σ 𝐶𝑜𝑟𝑟𝑒𝑐𝑡𝑆𝑒𝑔𝑚𝑒𝑛𝑡ሾ𝑗ሿ.𝐷𝑢𝑟𝑎𝑡𝑖𝑜𝑛𝑚𝑗=0σ 𝑆𝑒𝑔𝑚𝑒𝑛𝑡 ሾ𝑖ሿ.𝐷𝑢𝑟𝑎𝑡𝑖𝑜𝑛𝑁𝑖=0

N: the total number of the segments after beingpartitioned by a segmentation method.m: # of segments our approach correctly predicted

17

Experiment: Result

00.10.20.30.40.50.60.70.80.9

Decision Tree

SVM Bayes net CRF

Acc

urac

y

Inference Model

Accuracy by LengthAccuracy by Duration

Inferring accuracy of transportation mode over change point-based segmentation method

• Inference performance

18

Experiment

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

50 100 150 200 250 300

Reca

ll

Distance (m)

Decision Tree

SVM

Bayes net

CRF

Recall of change point using change point based

segmentation method

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

0.55

50 100 150 200 250 300

Prec

ision

Distance (m)

Decision Tree

SVM

Bayes net

CRF

Precision of change point using change point based segmentation method

• Inference performance of change point

19

Experiment: Result

change point uniform duration

(120 s)uniform length

(100 m)

Accuracy by Length 0.685 0.647 0.399

Accuracy by Duration 0.753 0.701 0.674

Recall/change point 0.887 0.867 0.867

Precision/change point 0.406 0.197 0.148

Comparison of different segmentation methods using Decision Tree

20

Experiment: Result

Comparison of inference results of CRF over different segmentation methods

change point

uniform duration(90 s)

uniform length(150 m)

Accuracy by Length 0.528 0.524 0.617

Accuracy by Duration 0.358 0.413 0.525

Recall/ change point 0.281 0.121 0.656

Precision /change point 0.286 0.070 0.159

21

Conclusion

Change Point based

Uniform Duration

based

Uniform Length based

SVM

Bayesian Net

Decision Tree

CRF

Segmentation method

Inference method

22

Future work

• Identify more valuable features• Location-constraint conditional probability• Improving prediction performance of CRF-based approach