inferring travel purpose from crowd-augmented human mobility data
DESCRIPTION
Talk slides for Urb-IoT conference 2014TRANSCRIPT
Inferring Travel Purpose from Crowd-Augmented
Human Mobility Data
!Zack Zhu, Ulf Blanke, and Gerhard Tröster
Wearable Computing Lab @ ETH Zurich !
The First International Conference on IoT in Urban Space, Rome, Italy !
October 27, 2014
Urban Travel
*Gibson, J. J. The theory of affordances. Hilldale, USA (1977)
Affordance*“The affordances of the environment are what it offers the animal, what it
provides or furnishes, either for good or ill.”
*Gibson, J. J. The theory of affordances. Hilldale, USA (1977)
Affordance*(Urban)“The affordances of the environment are what it offers the animal, what it
provides or furnishes, either for good or ill.”
*Gibson, J. J. The theory of affordances. Hilldale, USA (1977)
Affordance*(Urban)“The affordances of the environment are what it offers the animal, what it
provides or furnishes, either for good or ill.”
*Gibson, J. J. The theory of affordances. Hilldale, USA (1977)
Affordance*(Urban)“The affordances of the environment are what it offers the animal, what it
provides or furnishes, either for good or ill.”
*Gibson, J. J. The theory of affordances. Hilldale, USA (1977)
Affordance*(Urban)“The affordances of the environment are what it offers the animal, what it
provides or furnishes, either for good or ill.”
Why do people travel?
Why do people travel?
WHO?
WHEN?
WHERE? (what’s-there)
Why do people travel?
WHO?
WHEN?
What is at the origin and what is at the destination?
When did we depart and arrive? How long did we stay?
What is my age? What do I do?
WHERE? (what’s-there)
Why do people travel?
WHO?
WHEN?
What is at the origin and what is at the destination?
When did we depart and arrive? How long did we stay?
What is my age? What do I do?
Given who, when, where, can we infer why?
WHERE? (what’s-there)
Puget Sound Research Council Travel Survey*
•10,372 individuals •87,000 trips •48-hour survey period •Geo-coordinates + travel
purpose class
* Puget Sound Regional Council. 2006 Household Activity Survey: General User’s Guide. http://www.psrc.org/assets/4358/2006HHSurveyUserGuide.pdf, 2006.
Puget Sound Research Council Travel Survey*
•10,372 individuals •87,000 trips •48-hour survey period •Geo-coordinates + travel
purpose class
* Puget Sound Regional Council. 2006 Household Activity Survey: General User’s Guide. http://www.psrc.org/assets/4358/2006HHSurveyUserGuide.pdf, 2006.
CLASSES INSTANCESHOME 37,122WORK 11,776
PERSONAL BUSINESS 7,877
ACCOMPANYING OTHERS 7,819
SHOPPING 7,524RECREATION 4,766EATING OUT 4,041ATTENDING
SCHOOL 3,468
SOCIAL 3,202
Work School Social RecreationHome
Students Workers
Work School Social RecreationHome
Students Workers
Modelling Procedure• Feature-level fusion of who,
when, where
• Interpersonal 10-fold cross-validation
• L1-regularized Linear SVM
• Suitable for high-dimensional features spaces
• Feature coefficient magnitude useful for model interpretation
Who When Where Label
Why do people travel?
WHO?
WHEN?
WHERE? What is at the origin and what is at the destination?
When did we depart and arrive? How long did we stay?
What is my age? What do I do?
Given who, when, where, can we infer why?
Feature ConstructionWHO?WHEN?
Feature ConstructionWHO?WHEN?
4:34 pm
5:23 pm
Feature ConstructionWHO?WHEN?
4:34 pm
5:23 pm
[…, 0, 1,1, 0, …]
Feature ConstructionWHO?WHEN?
4:34 pm
5:23 pm
6:00 pm
[…, 0, 1,1, 0, …]
Feature ConstructionWHO?WHEN?
4:34 pm
5:23 pm
6:00 pm
[…, 0, 1,1, 0, …]
47 min
0.0326 days
Feature ConstructionWHO?WHEN?
4:34 pm
5:23 pm
6:00 pm
[…, 0, 1,1, 0, …]
47 min
0.0326 days
AGE GROUP
Under 5
5-10
16-17
18-24
25-34
35-44
45-54
55-64
65-74
75-84
85+
OCCUPATION
FT Worker
PT Worker
Retired Non-Worker
Other Non-Worker
Adult Student
Grade School 16+
Child Age 5-15
Child Age 0-4
Feature ConstructionWHERE? Coordinates (X,Y) -> Coordinates (X,Y)
Feature ConstructionWHERE? Coordinates (X,Y) -> Coordinates (X,Y)
The , also known as the Flavian Amphitheatre is an elliptical amphitheatre
in the centre of the city of Rome, Italy.
Originally known as the Flavian Amphitheatre, was commisioned in AD 72 by Emperor Vespasian.
Train:
Test:
Colosseum or Coliseum
Feature ConstructionWHERE? Coordinates (X,Y) -> Coordinates (X,Y)
The , also known as the Flavian Amphitheatre is an elliptical amphitheatre
in the centre of the city of Rome, Italy.
Originally known as the Flavian Amphitheatre, was commisioned in AD 72 by Emperor Vespasian.
Train:
Test:
Colosseum or Coliseum
Feature ConstructionWHERE? Coordinates (X,Y) -> Coordinates (X,Y)
The , also known as the Flavian Amphitheatre is an elliptical amphitheatre
in the centre of the city of Rome, Italy.Train:
Test:
Colosseum or Coliseum
Colosseum or Coliseum
WHERE?
Feature Constructionembedding (X,Y) -> embedding (X,Y)
WHERE?
Feature Construction
Foursquare as platform for semantic augmentation
embedding (X,Y) -> embedding (X,Y)
WHERE?
583 low-level venue semantics in 10 high-level categories
Feature Constructionembedding (X,Y) -> embedding (X,Y)
WHERE?
Rich, crowd-generated textual features
Feature Constructionembedding (X,Y) -> embedding (X,Y)
Feature Construction
Feature Construction
Cafe Business
[ 4 … … … … 7 …]
Feature Construction
Cafe Business
[ 4 … … … … 7 …]
best coffee in town!
bustling street great for people watching
try the salted caramel, bliss.
[ 1.23 … 0.24 … 0.32 …]
Tf-Idf weighted n-grams
What is “nearby”?
What is “nearby”?
Which aspect is important for inferring why?
42.38%'
50.09%'
66.23%'
75.28%'72.54%'
69.27%'
55.56%'
0.00%'
10.00%'
20.00%'
30.00%'
40.00%'
50.00%'
60.00%'
70.00%'
80.00%'
90.00%'
100.00%'
Demographical' Temporal' Spa<al' Fused' Without'Demographical'
Without'Temporal'
Without'Spa<al'
Average'Tes*ng'Accuracy'
Accuracy'Comparison'for'Combina*ons'of'Feature'Sets'
Which aspect is important for inferring why?
42.38%'
50.09%'
66.23%'
75.28%'72.54%'
69.27%'
55.56%'
0.00%'
10.00%'
20.00%'
30.00%'
40.00%'
50.00%'
60.00%'
70.00%'
80.00%'
90.00%'
100.00%'
Demographical' Temporal' Spa<al' Fused' Without'Demographical'
Without'Temporal'
Without'Spa<al'
Average'Tes*ng'Accuracy'
Accuracy'Comparison'for'Combina*ons'of'Feature'Sets'
Which aspect is important for inferring why?
42.38%'
50.09%'
66.23%'
75.28%'72.54%'
69.27%'
55.56%'
0.00%'
10.00%'
20.00%'
30.00%'
40.00%'
50.00%'
60.00%'
70.00%'
80.00%'
90.00%'
100.00%'
Demographical' Temporal' Spa<al' Fused' Without'Demographical'
Without'Temporal'
Without'Spa<al'
Average'Tes*ng'Accuracy'
Accuracy'Comparison'for'Combina*ons'of'Feature'Sets'
-19.72%
Which aspect is important for inferring why?
42.38%'
50.09%'
66.23%'
75.28%'72.54%'
69.27%'
55.56%'
0.00%'
10.00%'
20.00%'
30.00%'
40.00%'
50.00%'
60.00%'
70.00%'
80.00%'
90.00%'
100.00%'
Demographical' Temporal' Spa<al' Fused' Without'Demographical'
Without'Temporal'
Without'Spa<al'
Average'Tes*ng'Accuracy'
Accuracy'Comparison'for'Combina*ons'of'Feature'Sets'
Crowd-augmented “where”
0"
0.1"
0.2"
0.3"
0.4"
0.5"
0.6"
0.7"
0.8"
0.9"
1"
Accompanying"Others"
A=ending"School"
EaBng"Out" Home" Personal"Business"
RecreaBon" Shopping" Social" Work"
F1#Score)
F1#Score)Comparison)for)Using)Various)Spa6al)Features)
Venue"Category"L"High"Level"
Venue"Category"L"Low"Level"
Venue"Tips"
All"SpaBal"InformaBon"
Crowd-augmented “where”
0"
0.1"
0.2"
0.3"
0.4"
0.5"
0.6"
0.7"
0.8"
0.9"
1"
Accompanying"Others"
A=ending"School"
EaBng"Out" Home" Personal"Business"
RecreaBon" Shopping" Social" Work"
F1#Score)
F1#Score)Comparison)for)Using)Various)Spa6al)Features)
Venue"Category"L"High"Level"
Venue"Category"L"Low"Level"
Venue"Tips"
All"SpaBal"InformaBon"
Crowd-augmented “where”
0"
0.1"
0.2"
0.3"
0.4"
0.5"
0.6"
0.7"
0.8"
0.9"
1"
Accompanying"Others"
A=ending"School"
EaBng"Out" Home" Personal"Business"
RecreaBon" Shopping" Social" Work"
F1#Score)
F1#Score)Comparison)for)Using)Various)Spa6al)Features)
Venue"Category"L"High"Level"
Venue"Category"L"Low"Level"
Venue"Tips"
All"SpaBal"InformaBon"
Eating Out
Socializing
Attend School
Conclusion & Future Work• Augmentation of physical-world coordinates with
virtual-world contextualization cues • Travel inference with 75.28% accuracy (9 classes) • Feature importance: “where” > “when” > “who” • Model-driven indication of purpose context
• Feature engineering of “nearby” to consider travel time and venue importance weighting
• Validate model with real-time inference field study
Conclusion & Future Work• Augmentation of physical-world coordinates with
virtual-world contextualization cues • Travel inference with 75.28% accuracy (9 classes) • Feature importance: “where” > “when” > “who” • Model-driven indication of purpose context
• Feature engineering of “nearby” to consider travel time and venue importance weighting
• Validate model with real-time inference field study
Zack Zhu, Wearable Computing Lab, ETH Zurich [email protected]
What is “nearby”?
We investigate two notions of “nearby” for feature construction:
What is “nearby”?
We investigate two notions of “nearby” for feature construction:
Physical Radius
What is “nearby”?
We investigate two notions of “nearby” for feature construction:
Physical Radius
Rank Distance
What is “nearby”?
What is “nearby”?
What is “nearby”?
What is “nearby”?
What is “nearby”?
What is “nearby”?
Rank distance is adaptive against varying data density
What is “nearby”?
68.51%'
75.28%' 74.95%'73.83%' 73.11%' 72.79%'
58.93%'
65.92%'
71.89%'
72.74%' 72.52%'71.53%' 70.67%'
0' 100' 200' 300' 400' 500' 600' 700' 800' 900' 1000'
50.00%'
55.00%'
60.00%'
65.00%'
70.00%'
75.00%'
80.00%'
85.00%'
90.00%'
95.00%'
100.00%'
0' 20' 40' 60' 80' 100'
Physical)Distance)Radius)in)Meters)
Tes4ng)Accuracy)
Rank)Distance)Radius)in)Nearest)Neighbours)
Rank'Distance'
Physical'Distance'
Rank distance is adaptive against varying data density
Crowd-augmented “where”
Feature Rank
Feat
ure
Gro
ups
Demographical Features
Temporal Features
Venue Tips
Venue Category−Low Level
Venue Category−High Level
0 2000 4000 6000
●
●
●
●
●
o o o
o ooooo ooooooooooo
o
Accompanying Others
●
●
●
●
●o
Attending School
0 2000 4000 6000
●
●
●
●
●
o o o
oooooooooo
Eating Out
Demographical Features
Temporal Features
Venue Tips
Venue Category−Low Level
Venue Category−High Level
●
●
●
●
●
o
o
oo
Home
●
●
●
●
●
o o o o o o
ooo oooooooo
o
Personal Business
●
●
●
●
●
oo
o oo oo
oo ooooo oooooooooo
Recreation
Demographical Features
Temporal Features
Venue Tips
Venue Category−Low Level
Venue Category−High Level
●
●
●
●
●
o o o
ooooooooo
o
Shopping
0 2000 4000 6000
●
●
●
●
●
o o
oo o o oooooo
o o
Social
●
●
●
●
●
o
ooo oooooo
ooo
Work
Crowd-augmented “where”
Feature Rank
Feat
ure
Gro
ups
Demographical Features
Temporal Features
Venue Tips
Venue Category−Low Level
Venue Category−High Level
0 2000 4000 6000
●
●
●
●
●
o o o
o ooooo ooooooooooo
o
Accompanying Others
●
●
●
●
●o
Attending School
0 2000 4000 6000
●
●
●
●
●
o o o
oooooooooo
Eating Out
Demographical Features
Temporal Features
Venue Tips
Venue Category−Low Level
Venue Category−High Level
●
●
●
●
●
o
o
oo
Home
●
●
●
●
●
o o o o o o
ooo oooooooo
o
Personal Business
●
●
●
●
●
oo
o oo oo
oo ooooo oooooooooo
Recreation
Demographical Features
Temporal Features
Venue Tips
Venue Category−Low Level
Venue Category−High Level
●
●
●
●
●
o o o
ooooooooo
o
Shopping
0 2000 4000 6000
●
●
●
●
●
o o
oo o o oooooo
o o
Social
●
●
●
●
●
o
ooo oooooo
ooo
Work
Venue tips consistently rank with the highest (median) feature importance