This document summarizes research using smartphone sensors and data to analyze daily behaviors and activities. It presents three case studies: (1) using WiFi and GPS data to analyze public transportation routes and times; (2) correlating accelerometer and survey data to understand relationships between physical activity and subjective wellbeing; (3) the potential for behavioral interventions using sensor data. It also discusses challenges in working with large-scale smartphone data and opportunities for multidisciplinary research impact.

1. Analysing Daily Behaviours with Large-Scale
Smartphone Data
@neal_lathia
Computer Laboratory
University of Cambridge

2. Background
Smartphones as Research Tools
Case 1: Public Transport
Case 2: Subjective Wellbeing & Behaviour
Case 3: Behavioural Intervention
Challenges, Opportunities, Questions

3. Background

5. Smartphones as Research Tools

6. “by 2025, when most of today’s psychology
undergraduates will be in their mid-30s, more
than 5 billion people on our planet will be using
ultra-broadband, sensor-rich smartphones far
beyond the abilities of today’s iPhones, Androids,
and Blackberries.”
Miller

9. Accelerometer
GPS / Wi-Fi
Gyroscope
Bluetooth
Microphone
Humidity
Temperature
Phone / Text Logs
Device Logs
Social Media APIs
App Usage

10. Accelerometer | Physical Activity
GPS / Wi-Fi | Mobility
Gyroscope | Orientation
Bluetooth | Co-Location
Microphone | Ambient Audio
Humidity | Environment
Temperature | Environment
Phone / Text Logs | Socialising
Device Logs | Network
Social Media APIs | Socialising
App Usage | /Information Needs

11. Case 1: Public Transport
N. Lathia, L. Capra. Tube Star: Crowd-Sourced Experiences on Public
Transport. In 11th
International Conference on Mobile and Ubiquitous
Systems. London, December 2014.

14. Conclusion: potential for smartphones as a near-
real time passenger surveying tool to collect
qualitative trip data.
Major Limitation: amount of data received. Still
relies on reported behaviour.

16. Accelerometer (activity)
GPS / Wi-Fi (location)
Gyroscope (orientation)
Bluetooth (co-location)
Microphone (audio context)
Environment (temperature)
Phone / Text Logs (sociability)
Device Logs (e.g., network)
Social Media APIs (crowdsourcing)
App Usage (information seeking)

17. Method
1. Collect Wi-Fi scans that match
“Virgin Media Wi-Fi”
2. Manually label “unknown”
stations (“Where are you?”)
3. Apply heuristic-based clustering
algorithm to determine station
visits, paths, travel times.
Preliminary Data
34 users; 234,769 Wi-Fi scans,
106,793

18. Sequences of Wi-Fi Connections

19. Oyster Card transaction

20. Journey Planner:
Victoria Line to Oxford Circus
Central Line to Mile End
27 Minutes

21. Measured:
Victoria Line to Victoria
Circle/District to Mile End
29 minutes
21:54:02
21:56:06
21:58:13
22:12:00
22:15:41
22:23:05

22. Oyster card transaction

23. Journey Planner:
Piccadilly Line to Holborn
Central Line to West Ruislip
56 Minutes

24. Measured:
Piccadilly Line to King's Cross
Victoria Line to Oxford Circus
Central Line to West Ruislip
74 minutes
10:56:06
11:15:39
11:32:55
12:10:09

25. Capturing Routes: Given an O-D pair, count the % of times that
another station appears as an intermediary
Observing Mistakes? E.g., 2.18% of trips from Pimlico to Victoria
Station go via Green Park (wrong direction).
Non-Adjacent Pairs of Wi-Fi Connections

27. 40%
10%
10%
5%

29. 5%
41.70%
64.7%

30. Continued (lessened) limitations:
Data is not “complete” - phones do not always connect.
Data is now “noisy” by capturing route errors, “strange”
behaviours.
Direct application:
Transport route choices in individuals
With more scale:
Granular origin-destination + distributions of route data.
With more data:
I.e., precise locations of Wi-Fi hotspots (e.g., platform,
entrance)
With more sensors:
What actual behaviours are occurring?

31. Case 2: Subjective Wellbeing & Behaviour
N. Lathia, K. Rachuri, C. Mascolo, P. Rentfrow. Contextual Dissonance:
Design Bias in Sensor-Based Experience Sampling Methods. In ACM
International Joint Conference on Pervasive and Ubiquitous Computing.
Zurich, Switzerland. September 2013.
N. Lathia, G. Sandstrom, P. Rentfrow, C. Mascolo. Happy People Live
Active Lives. In prep.

32. “A sample of 222 undergraduates was screened
for high happiness using multiple confirming
assessment filters. We compared the upper 10%
of consistently very happy people with average
and very unhappy people. The very happy people
were highly social, and had stronger romantic and
other social relationships than less happy
groups...”
Diener, Seligman. Very Happy People. In Psychological Science 13 (1). Jan 2002.

37. angry anxious lonely
relaxedenthusiasticcalm

40. Hemminki, Nurmi, Tarkoma. Accelerometer-Based Transportation Mode Detection on
Smartphones. In ACM Sensys 2013.
Statistical: mean, standard deviation, median, etc.
Time: auto-correlation, mean-crossing rate, etc.
Frequency: FFT, spectral energy, etc.
Peak: volume, intensity, skewness, etc.
Segment: e.g., velocity change rate

41. Example: 85 Users

47. Case 3: Behavioural Intervention

51. Challenges & Questions

52. 1. Software Engineering / Expectations
2. Marketing
3. Control over target population
4. Understanding sensor data
5. Writing code
6. Finding research value

53. 1. Blurred lines between research and practice
2. High potential for multi-disciplinary impact
3. Cheap to roll-out to huge audiences
4. Accessible to 'everyone'
5. Wearables are coming!

54. Can I run a study like
Emotion Sense?
Yes, with Easy M. A
generalised sensor-
enhanced experience
sampling tool.

55. @neal_lathia
neal.lathia@cl.cam.ac.uk
http://www.cl.cam.ac.uk/~nkl25/