Presentation I used while defending my thesis on MEILI: Multiple Day Travel Behaviour Data Collection, Automation and Analysis. Thesis available at: http://kth.diva-portal.org/smash/record.jsf?pid=diva2%3A1204245&dswid=7962

1. MEILI
Multiple Day Travel Behaviour Data Collection,
Automation and Analysis
Adrian C. Prelipcean
Dept. of Transport Science
KTH Royal Institute of Technology
Stockholm, Sweden
acpr@kth.se
@Adi Prelipcean
adrianprelipcean.github.io
5 June 2018

2. Overview
This presentation will be about:
1. Introduction and short history of travel diaries
2. Thesis main objectives
3. Methodology
– Travel diary collection and automation
– Multiple source travel diary comparison
– Sequential analysis of multiple day travel diaries
4. Finishing Words
– Addressing current research needs
– Future research directions
2

3. Travel behaviour
How do we use travel behaviour?
Some of the main reasons for analyzing travel behaviour are:
to investigate the reasons and mechanisms that underlie
an individual’s travel decision making process,
3

4. Travel behaviour
How do we use travel behaviour?
Some of the main reasons for analyzing travel behaviour are:
to investigate the reasons and mechanisms that underlie
an individual’s travel decision making process,
to predict the effect of implementing new transportation
policies or changing the transportation infrastructure, or
3

5. Travel behaviour
How do we use travel behaviour?
Some of the main reasons for analyzing travel behaviour are:
to investigate the reasons and mechanisms that underlie
an individual’s travel decision making process,
to predict the effect of implementing new transportation
policies or changing the transportation infrastructure, or
to understand the dynamic of transportation movement
within study areas.
3

6. Travel behaviour
How do we get insights into travel behaviour?
Travel Diaries - a way of summarizing where, why and how a
user traveled during a defined time frame by specifying:
The destination of a trip
Img: http://soarministries.com/hp_wordpress/wp-content/uploads/2011/08/Destinations-Icon.jpg 4

7. Travel behaviour
How do we get insights into travel behaviour?
Travel Diaries - a way of summarizing where, why and how a
user traveled during a defined time frame by specifying:
The destination of a trip
The trip’s purpose
Img: https://cdn2.vox-cdn.com/thumbor/93Yaxs7y3Tb8tzFfppyRsSn_yN8=/1020x0/cdn0.vox-cdn.com/ 4

8. Travel behaviour
How do we get insights into travel behaviour?
Travel Diaries - a way of summarizing where, why and how a
user traveled during a defined time frame by specifying:
The destination of a trip
The trip’s purpose
The means of transportation, i.e., trip legs
Img: https://d3ui957tjb5bqd.cloudfront.net/images/screenshots/products/4/42/42990/ 4

9. Travel diaries
Short history
(1960s) Travel diaries are commonly used to better
understand travel behaviour.
5

10. Travel diaries
Short history
(1960s) Travel diaries are commonly used to better
understand travel behaviour.
(1980s) Scientists studied the effect of changing the
focus of questions from destinations to activities and
found an increase in the response rate when using
activity travel diaries
5

11. Travel diaries
Short history
(1960s) Travel diaries are commonly used to better
understand travel behaviour.
(1980s) Scientists studied the effect of changing the focus
of questions from destinations to activities
(late 1980s) Researchers started studying multiple day
travel behaviour but the high costs of collecting travel
diaries over multiple days prohibit the new methods’
wide use
5

12. Travel diaries
Short history
(1960s) Travel diaries are commonly used to better
understand travel behaviour.
(1980s) Scientists studied the effect of changing the focus
of questions from destinations to activities
(late 1980s) Researchers started studying multiple day
travel behaviour
(late 1990s) Countries have started seeing a decrease in
the response rate to travel diaries
5

13. Travel diaries
Short history
(1960s) Travel diaries are commonly used to better
understand travel behaviour.
(1980s) Scientists studied the effect of changing the focus
of questions from destinations to activities
(late 1980s) Researchers started studying multiple day
travel behaviour
(late 1990s) Countries have started seeing a decrease in
the response rate to travel diaries
(2000) Scientists started collecting data with
dedicated GPS devices from cars or pedestrians to
extract similar data as processed travel diaries
5

14. Travel diaries
Short history
(1960s) Travel diaries are commonly used to better
understand travel behaviour.
(1980s) Scientists studied the effect of changing the focus
of questions from destinations to activities
(late 1980s) Researchers started studying multiple day
travel behaviour
(late 1990s) Countries have started seeing a decrease in
the response rate to travel diaries
(2000) Scientists started collecting data with dedicated
GPS devices from cars or pedestrians
(2010) Smartphones became popular devices with a
high penetration rate (1 billion devices world wide
reached in 2012 Q3)
5

15. Travel diaries
Short history
(1960s) Travel diaries are commonly used to better
understand travel behaviour.
(1980s) Scientists studied the effect of changing the focus
of questions from destinations to activities
(late 1980s) Researchers started studying multiple day
travel behaviour
(late 1990s) Countries have started seeing a decrease in
the response rate to travel diaries
(2000) Scientists started collecting data with dedicated
GPS devices from cars or pedestrians
(2010) Smartphones became popular devices
(Present) Scientists experiment with segmenting
trajectories into trips and triplegs and detecting the
purpose and destination of trips, and travel modes of
triplegs 5

16. Thesis Main Objectives
Research objectives for this thesis
1. Collect multiple day travel diaries from the same users at
a low cost by using smartphones to collect trajectories
and classifiers to transform trajectories into travel diaries
(Papers I - IV, VIII)
2. Compare the new methods with the classical ways of
collecting travel diaries, e.g., filling in forms, phone
interviews, etc.(Papers V-VII)
3. Continue the current research on multiple day travel
behaviour, with a focus on sequences (Paper X)
4. Frame the findings and progress within the current state
of the art and the research needs (Paper VIII, IX)
6

17. Designing a system to collect travel diaries
A trajectory is a sequence of
timestamped latitude longitude pairs
7

18. Designing a system to collect travel diaries
A trajectory is a sequence of
timestamped latitude longitude pairs
A trip contains the part of a trajectory
travelled by the user on the way to
perform an activity at a destination
7

19. Designing a system to collect travel diaries
A trajectory is a sequence of
timestamped latitude longitude pairs
A trip contains the part of a trajectory
travelled by the user on the way to
perform an activity at a destination
A tripleg contains the part of a
trajectory belonging to a trip that is
traveled by the same travel mode
7

20. Designing a system to collect travel diaries
Questions
1. How to collect trajectories? (Paper I)
2. How to store trajectories and travel diaries? (Paper II)
3. How to transform the collected trajectories into travel
diaries? (Paper III, IV, V)
8

21. Collecting trajectories
Motivation
Smartphones have a high penetration
rate across different regions and
demographics
Img: https://www.statista.com/statistics/488351/smartphone-penetration-sweden/
9

22. Collecting trajectories
Motivation
Smartphones have a high penetration
rate across different regions and
demographics
Users see smartphones as personal
devices
Img: https://www.statista.com/statistics/488351/smartphone-penetration-sweden/
9

23. Collecting trajectories
Motivation
Smartphones have a high penetration
rate across different regions and
demographics
Users see smartphones as personal
devices
Smartphones have a multitude of
sensors available (including GPS
receivers), however collecting data
comes with unavoidable battery
consumption
Img: http://www.vensi.com/wp-content/uploads/2017/06/Sensor-Technology-in-Smartphones.jpg
9

24. Collecting trajectories
MEILI Mobility Collector
MEILI Mobility Collector
– Equidistance and equitime sampling of
locations
– GPS and accelerometer data fusion
10

25. Collecting trajectories
MEILI Mobility Collector
MEILI Mobility Collector
– Equidistance and equitime sampling of
locations
– GPS and accelerometer data fusion
Point- and period-based annotations
10

26. Collecting trajectories
MEILI Mobility Collector
MEILI Mobility Collector
– Equidistance and equitime sampling of
locations
– GPS and accelerometer data fusion
Point- and period-based annotations
Battery efficient collection based on:
– in-doors non-movement (acc.)
– equidistance sampling (speed)
10

27. Collecting trajectories
MEILI Mobility Collector
MEILI Mobility Collector
– Equidistance and equitime sampling of
locations
– GPS and accelerometer data fusion
Point- and period-based annotations
Battery efficient collection based on:
– in-doors non-movement (acc.)
– equidistance sampling (speed)
Achieved reasonable user satisfaction
for battery consumption while
collecting data
10

28. Storing trajectories and travel diaries
The only data model available for travel diaries is either a
byproduct of the used forms for asking travel questions or
a byproduct of the software used for analyzing travel
diaries
11

29. Storing trajectories and travel diaries
Proposed a data model for storing travel diaries and
linking the travel diaries to collected trajectories
11

30. Storing trajectories and travel diaries
11

31. Storing trajectories and travel diaries
11

32. Generating travel diaries from trajectories
Overview of state of the art
12

33. Generating travel diaries from trajectories
Confusion for applying state of the art
12

34. Generating travel diaries from trajectories
More confusion for applying state of the art
12

35. Generating travel diaries from trajectories
Q: What do the declared precision / accuracy values actually
mean?
12

36. Generating travel diaries from trajectories
Q: What do the declared precision / accuracy values actually
mean?
A: Usually, the percentage of instances when an element is
correctly inferred given that all the previously performed
operations are 100% accurate (e.g., trip segmentation before
purpose inference)
12

37. Generating travel diaries from trajectories
Q: What do the declared precision / accuracy values actually
mean?
A: Usually, the percentage of instances when an element is
correctly inferred given that all the previously performed
operations are 100% accurate (e.g., trip segmentation before
purpose inference)
Addendum: However, the precision / accuracy values are
computed differently for different disciplines (Paper IV)
12

38. Different ways of looking at travel
Different fields that have different views on
travel:
Transport Science - Q: How were users
travelling during a defined period?
13

39. Different ways of looking at travel
Different fields that have different views on
travel:
Transport Science - Q: How were users
travelling during a defined period?
Location Based Services - Q: How is a
user travelling now?
13

40. Different ways of looking at travel
Different fields that have different views on
travel:
Transport Science - Q: How were users
travelling during a defined period?
Location Based Services - Q: How is a
user travelling now?
Human Geography - Q: How can a
trajectory be segmented into parts
that can be enriched with domain
specific semantics?
13

41. Different ways of looking at travel
Different fields that have different views on
travel:
Transport Science - Q: How were users
travelling during a defined period?
Location Based Services - Q: How is a
user travelling now?
Human Geography - Q: How can a
trajectory be segmented into parts
that can be enriched with domain
specific semantics?
Every domain has a unique and
non-transferable definition of error
13

42. Measuring the performance of trajectory
segmentation
In transport science, trips are matched based on temporal
(or spatio-temporal) overlap of the start and end of the
trip trajectory
14

43. Measuring the performance of trajectory
segmentation
In transport science, trips are matched based on temporal
(or spatio-temporal) overlap of the start and end of the
trip trajectory
14

44. Measuring the performance of trajectory
segmentation
In transport science, trips are matched based on temporal
(or spatio-temporal) overlap of the start and end of the
trip trajectory
Q: What happens when a system misses a trip?
A: The system fails to capture the semantics of one trip and
affects at least one of its neighbours.
14

45. Robust errors for trajectory segmentations
Method based on Allen (1983) -
Maintaining knowledge about
temporal intervals
15

46. Robust errors for trajectory segmentations
Method based on Allen (1983) -
Maintaining knowledge about
temporal intervals
Replaced rigid interval matching with
penalty-based interval alignment
15

47. Robust errors for trajectory segmentations
Method based on Allen (1983) -
Maintaining knowledge about
temporal intervals
Replaced rigid interval matching with
penalty-based interval alignment
Introduced new trajectory
segmentation performance metrics :
– Precision and Recall
– Shift-in and Shift-out penalties
– Oversegmentation
15

48. Robust errors for trajectory segmentations
Method based on Allen (1983) -
Maintaining knowledge about
temporal intervals
Replaced rigid interval matching with
penalty-based interval alignment
Introduced new trajectory
segmentation performance metrics :
– Precision and Recall
– Shift-in and Shift-out penalties
– Oversegmentation
Can compare the achievable precision
given a segmentation algorithm and
the achieved precision given a travel
mode classifier
15

49. Robust errors for trajectory segmentations
Precision
Shift-Out Space
Method based on Allen (1983) -
Maintaining knowledge about
temporal intervals
Replaced rigid interval matching with
penalty-based interval alignment
Introduced new trajectory
segmentation performance metrics :
– Precision and Recall
– Shift-in and Shift-out penalties
– Oversegmentation
Can compare the achievable precision
given a segmentation algorithm and
the achieved precision given a travel
mode classifier
Can compare different travel mode
classifiers
15

50. Robust errors for trajectory segmentations
Method based on Allen (1983) -
Maintaining knowledge about
temporal intervals
Replaced rigid interval matching with
penalty-based interval alignment
Introduced new trajectory
segmentation performance metrics
Can compare the achievable precision
given a segmentation algorithm and
the achieved precision given a travel
mode classifier
Can compare different travel mode
classifiers
Proposed three mode detection
strategies:
– Implicit segmentation
– Explicit-holistic segmentation
– Explicit-consensus based segmentation15

51. Generating travel diaries from trajectories
A new research objective
While there is a clear distinction in between different
methods’ performance, none of the tested methods are
ready for the full automation of travel diaries
16

52. Generating travel diaries from trajectories
A new research objective
How to design a system that can transform trajectories into
travel diaries, allows for users to correct them, and learns from
the user annotations? (Paper II)
16

53. MEILI
Architecture of a self-improving system for collecting travel
diaries
17

54. MEILI
Architecture of a self-improving system for collecting travel
diaries
Precision and recall over time (15 travel modes)
17

55. MEILI
Making the system readily available for multiple region travel
diary collection
18

56. State of the art travel diary collection systems
Main types of travel diary collection systems
19

57. State of the art travel diary collection systems
Main travel diary collection systems
20

58. State of the art travel diary collection systems
Main travel diary collection systems
20

59. Comparing different travel diary collection systems
21

60. Comparing different travel diary collection systems
Trips best matched on time and
purpose - geocoding is error prone
22

61. Comparing different travel diary collection systems
Trips best matched on time and
purpose - geocoding is error prone
Types of captured information:
– Intrinsic - how well an entity, incl. its
attributes, is captured?
– Extrinsic - how well do different
systems agree on the capture of an
entity?
22

62. Comparing different travel diary collection systems
Trips best matched on time and
purpose - geocoding is error prone
Types of captured information:
– Intrinsic - how well an entity, incl. its
attributes, is captured?
– Extrinsic - how well do different
systems agree on the capture of an
entity?
Spatial and temporal indicators
measure intrinsic information
22

63. Comparing different travel diary collection systems
Trips best matched on time and
purpose - geocoding is error prone
Types of captured information:
– Intrinsic - how well an entity, incl. its
attributes, is captured?
– Extrinsic - how well do different
systems agree on the capture of an
entity?
Spatial and temporal indicators
measure intrinsic information
Intrinsic and extrinsic information
useful for in-depth analysis
22

64. Comparing different travel diary collection systems
Trips best matched on time and
purpose - geocoding is error prone
Types of captured information:
– Intrinsic - how well an entity, incl. its
attributes, is captured?
– Extrinsic - how well do different
systems agree on the capture of an
entity?
Spatial and temporal indicators
measure intrinsic information
Intrinsic and extrinsic information
useful for in-depth analysis
Unifying framework of previous
concepts
22

65. Case studies with MEILI
Location: Stockholm
Number of participants: 11 (first case
study), 30 (second case study), 171
(third case study)
23

66. Case studies with MEILI
Location: Stockholm
Number of participants: 11 (first case
study), 30 (second case study), 171
(third case study)
The overall descriptive statistics for
both systems are similar.
23

67. Case studies with MEILI
Location: Stockholm
Number of participants: 11 (first case
study), 30 (second case study), 171
(third case study)
The overall descriptive statistics for
both systems are similar.
The percentage of trips captured by
MEILI increased between the two case
studies.
23

68. Case studies with MEILI
Location: Stockholm
Number of participants: 11 (first case
study), 30 (second case study), 171
(third case study)
The overall descriptive statistics for
both systems are similar.
The percentage of trips captured by
MEILI increased between the two case
studies.
The reasons for missing a trip
changed between the two case studies.
23

69. Case studies with MEILI
MEILI collects more trips and at a
finer granularity, but it does not
collect all trips
There is no clear superior method for
collecting travel diaries
24

70. Sequential stability of travel behaviour activities
25

71. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns
26

72. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns
26

73. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns but the complexity increases
with the number of days and activities
26

74. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns but the complexity increases
with the number of days and activities
We can extract the Longest Common
Subsequences to identify common
patterns between days and travelers
26

75. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns but the complexity increases
with the number of days and activities
We can extract the Longest Common
Subsequences to identify common
patterns between days and travelers
Half of activities are performed in the
same order in the user base
26

76. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns but the complexity increases
with the number of days and activities
We can extract the Longest Common
Subsequences to identify common
patterns between days and travelers
Half of activities are performed in the
same order in the user base
The travel mode scheduling is more
diverse, especially on weekends
26

77. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns but the complexity increases
with the number of days and activities
We can extract the Longest Common
Subsequences to identify common
patterns between days and travelers
Half of activities are performed in the
same order in the user base
The travel mode scheduling is more
diverse, especially on weekends
Habitual activities are seldom part of
inter-personal sequences
26

78. Sequential stability of travel behaviour activities
It is intuitive to notice simple schedule
patterns but the complexity increases
with the number of days and activities
We can extract the Longest Common
Subsequences to identify common
patterns between days and travelers
Half of activities are performed in the
same order in the user base
The travel mode scheduling is more
diverse, especially on weekends
Habitual activities are seldom part of
inter-personal sequences
Intra personal sequences contain more
habitual activities
26

79. Research priorities
What are the main worries of researchers in this field
Based on discussions with researchers (Paper IX):
Current status of travel diary collection systems based on
smartphones and GPS receivers
– Establishing ground truth - Paper V
– Assuring the sample representativeness by combining multiple data
collection methods (e.g., paper and pen declaration and smartphone
based collection)
– Minimizing the cost of maintenance and distribution of travel diary
collection system - Paper VIII
– The tradeoffs in cost for GPS receivers / loggers (high distribution cost)
and smartphones (high development cost)
27

80. Research priorities
What are the main worries of researchers in this field
Based on discussions with researchers (Paper IX):
Current status of travel diary collection systems based on
smartphones and GPS receivers
Methodologies to extract semantics from GPS trajectories
and auxiliary data
– Investigate relevant performance measures for travel diary collection
systems that take into account the sequential, spatial and temporal
distribution of GPS data - Paper III
– Implement classification methods that leverage the availability of
auxiliary data sources (e.g., POI data sets, transit feeds)
27

81. Research priorities
What are the main worries of researchers in this field
Based on discussions with researchers (Paper IX):
Current status of travel diary collection systems based on
smartphones and GPS receivers
Methodologies to extract semantics from GPS trajectories
and auxiliary data
Performance and usability considerations for using
smartphones and GPS receivers to collect data for generating
travel surveys
– Limit battery consumption for travel diary collection systems - Paper I
– Strategies for onboarding users to collect GPS trajectories with
smartphones - Paper VIII
– Leverage the varying accuracy of GPS records between multiple devices
– Improve user interface and user experience to allow for a wider
demographic use
27

82. Research priorities
What are the main worries of researchers in this field
Based on discussions with researchers (Paper IX):
Current status of travel diary collection systems based on
smartphones and GPS receivers
Methodologies to extract semantics from GPS trajectories
and auxiliary data
Performance and usability considerations for using
smartphones and GPS receivers to collect data for generating
travel surveys
Potential applications of finer spatio-temporal granularity
survey data
– Level of detail for data storage and data sharing between institutes to
preserve privacy
– Establishing an appropriate abstraction level for collected GPS data to
avoid redundant data collection - Paper I
– How to get valid informed consent from users
– Opportunity to include the collected GPS data for market research
27

83. Main Contributions
What are the main contributions of this thesis?
Data Collection
– MEILI Mobility Collector
– MEILI Travel Diary
– MEILI Artificial Intelligence
28

84. Main Contributions
What are the main contributions of this thesis?
Data Collection
– MEILI Mobility Collector
– MEILI Travel Diary
– MEILI Artificial Intelligence
Data Automation
– Interval alignment
– Travel mode (implicit, explicit-holistic, explicit-consensus), destination
and purpose inference
– New performance measures for trajectory segmentation
28

85. Main Contributions
What are the main contributions of this thesis?
Data Collection
– MEILI Mobility Collector
– MEILI Travel Diary
– MEILI Artificial Intelligence
Data Automation
– Interval alignment
– Travel mode (implicit, explicit-holistic, explicit-consensus), destination
and purpose inference
– New performance measures for trajectory segmentation
Data Analysis
– Trip matching methodology
– Spatial and temporal indicators
– Ground truth candidates
– Stability and variability of travel patterns
28

86. Future research directions
On MEILI
– The UI of MEILI has to be simplified and made bug-free to improve UX
and to reduce the drop-rate of users
– The machine learning part of MEILI needs to be improved before it can
be used in other real-world travel diary collection sessions
– Further standardize deployment and hosting
– Rewrite the algorithms for segmenting trajectories into trips and triplegs
to use self-learning principles similar to the other methods
– Identify the minimum period a user has to annotate for MEILI to reach a
sufficient performance when generating travel diaries
29

87. Future research directions
On MEILI
– The UI of MEILI has to be simplified and made bug-free to improve UX
and to reduce the drop-rate of users
– The machine learning part of MEILI needs to be improved before it can
be used in other real-world travel diary collection sessions
– Further standardize deployment and hosting
– Rewrite the algorithms for segmenting trajectories into trips and triplegs
to use self-learning principles similar to the other methods
– Identify the minimum period a user has to annotate for MEILI to reach a
sufficient performance when generating travel diaries
On general research
– The replacement of traditional travel diary collection systems should be
studied also in terms of collection bias, to observe which system can
collect unbiased data
– There is a need for a general benchmark dataset that scientists should
use for testing the automation of extracting different travel diary parts
– Uniform reporting of errors is needed in the scientific community
– Increase cooperation between research and industry
– Recheck the widely accepted assumptions derived from single day travel
diary studies with multiple day data
29

88. Thank you for your attention!
Questions and Discussions
Adrian C. Prelipcean
Phd Student
Transport Science
KTH, Royal Institute of Technology
http://adrianprelipcean.github.io/
acpr@kth.se