1. The document discusses using anonymized smart card data from public transport systems to analyze travel behaviors and develop personalized recommendations and predictions. 2. Models are able to accurately predict station disruptions, travel times, and recommend cheaper fare purchase options using only anonymized user data without direct information on schedules or networks. 3. Analyzing such data could enable powerful personalized information systems while also providing insights into policy impacts and community well-being.

1. @neal_lathia
computer laboratory: university of cambridge

2. online offline

3. urban
data mining
web
urbanmining.wordpress.com

7. online
user data + algorithms → relevance ☺

11. public transport
user data + algorithms → relevance

12. “smart” cards
1 facilitate payment
2 collect user data

13. “smart” cards
time-stamped locations,
modality, payments,
user categories
anonymised with
persistent user ids

14. “smart” cards datasets
100% - 1 month
~5.1 million people
~78.8 million trips
5% - 2 x 83 days
~300k people
~7.7 million trips

18. Purchase Geography Mobility Flow
45
Zone 1
PAYG Zone 2
40
Travel Cards Zone 3
35 Zone 4
Zone 5
30 Zone 6
25
20
15
10
5 arrive
0
1 2 3 4 5 6 7 8 9

19. using transport data for...
1 predicting disruption relevance
2 personalised travel time
3 fare purchase recommendation

21. can we use transport data for...
1 predicting disruption relevance
i.e., rank station importance correctly?

22. can we use transport data for...
predicting disruption relevance
i.e., rank station importance correctly?
(where you will go in the future)

23. percentile ranking
0.0 (best)
…
0.5 (random)
…
1.0 (inverse)

24. percentile ranking
0.0 (best)
...
0.25 (rank stations by popularity)
...
0.5 (random)
…
1.0 (inverse)

25. percentile ranking
0.0 (best)
...
0.06 (factor in user's history)
...
0.25 (rank stations by popularity)
...
0.5 (random)
…
1.0 (inverse)

26. percentile ranking
0.0 (best)
…
0.05 (“those who touch in here also touch in at...”)
...
0.06 (factor in user's history)
...
0.25 (rank stations by popularity)
...
0.5 (random)
…
1.0 (inverse)

27. accurate ranking without
1 explicitly asking
2 network topology, rail schedule

28. using transport data for...
1 predicting disruption relevance
2 personalised travel time

32. can we use transport data for...
2 predict your travel time
i.e., time between touch in/out?

33. mean absolute error (minutes)
0.0 (best)
…

34. mean absolute error (minutes)
0.0 (best)
…
9.82 (time tabled)

35. mean absolute error (minutes)
0.0 (best)
…
3.30 (mean time)
...
9.82 (time tabled)

36. mean absolute error (minutes)
0.0 (best)
…
3.28 (“people who travel at this time...”)
3.30 (mean time)
...
9.82 (time tabled)

37. mean absolute error (minutes)
0.0 (best)
…
3.17 (“people who are as familiar as you...”)
3.28 (“people who travel at this time...”)
3.30 (mean time)
...
9.82 (time tabled)

38. mean absolute error (minutes)
0.0 (best)
…
3.13 (“your trips in the past...”)
3.17 (“people who are as familiar as you...”)
3.28 (“people who travel at this time...”)
3.30 (mean time)
...
9.82 (time tabled)

39. accurate predictions without
1 explicitly asking
2 network topology, rail schedule
3 ongoing disruptions, delays

40. using transport data for...
1 predicting disruption relevance
2 personalised travel time
3 fare purchase recommendation

41. 30
Purchase Behaviour
Travel Cards
25
PAYG
20
% Purchases
15
10
5
0
Mon Tue Wed Thu Fri Sat Sun
45
Purchase Geography
Mobility Flow
40
PAYG Zone 1
Travel Cards Zone 2
35 arrive Zone 3
30 Zone 4
Zone 5
25 Zone 6
20
15
10
5
0
1 2 3 4 5 6 7 8 9

42. (a) high regularity in purchases & movements
(b) small increments, short terms
(c) purchase on refused entry?

43. are people making the right choice?

44. £200 million
overspend

45. (a) failure to predict your movements
(b) failing to match mobility with fares

46. can we use transport data for...
3 predict the fares you should buy
i.e., what will be cheapest?

47. classification accuracy
0.0% (worst)
...
100% (oracle)

48. classification accuracy
0.0 (worst)
…
77% everyone on pay as you go
...
100% (oracle)

49. classification accuracy
0.0 (worst)
…
77% everyone on pay as you go
80% naïve bayes
...
100% (oracle)

50. classification accuracy
0.0 (worst)
…
77% everyone on pay as you go
80% naïve bayes
…
97% (“people like you should have bought...”)
100% (oracle)

51. classification accuracy
0.0 (worst)
…
77% everyone on pay as you go
80% naïve bayes
…
97% (“people like you should have bought...”)
98% decision trees
100% (oracle)

52. money saved
£0.0 (worst)
…
£326,447.95 everyone on pay as you go
£393,585.81 naïve bayes
…
£465,822.17 (“people like you...”)
£473,918.38 decision trees
£479,583.91 (oracle)

53. “smart” cards
1 facilitate payment
2 collect user data
3 enable powerful,
personalised
information systems

55. using transport data for...
1 behaviours ~ policy & incentives
2 community well-being

56. References
N. Lathia, J. Froehlich, L. Capra. Mining Public Transport Usage for Personalised Intelligent
Transport Systems. In IEEE International Conference on Data Mining. December 2010, Sydney,
Australia.
N. Lathia, C. Smith, J. Froehlich, L. Capra. Individuals Among Commuters: Building
Personalised Transport Information Systems from Fare Collection Systems. Under submission.
N. Lathia, L. Capra. Mining Mobility Data to Minimise Travellers' Spending on Public Transport.
In ACM International Conference on Knowledge Discovery and Data Mining. August 2011. San
Diego, USA.
N. Lathia, L. Capra. How Smart is Your Smart Card? Measuring Travel Behaviours,
Perceptions, and Incentives. In ACM International Conference on Ubiquitous Computing.
September 2011. Beijing, China.
N. Lathia, D. Quercia, J. Crowcroft. The Hidden Image of the City: Sensing Community Well-
Being from Urban Mobility. To Appear, 10th International Conference on Pervasive Computing.
June 2012. Newcastle, UK.