From Network Noise to Social Signals

Transcription

1 From Network Noise to Social Signals NETWORK-SENSING FOR BEHAVIOURAL MODELLING IN PRIVATE AND SEMI-PUBLIC SPACES Afra Mashhadi Bell Labs, Nokia 23rd May

2 WHAT CAN BEHAVIOUR MODELLING BE USED FOR? n Space Analytics á People Analytics Spatio-Temporal Analysis Informed Space Management Personalised Space Recommendation Better Resource Management Crowd Flow Management Profiling of Individual and Groups Recommending behavioural intervention Extracting Communities and their Properties

3 WHERE CAN BEHAVIOUR MODELLING BE USED IN? Private Spaces: e.g., Workplaces Semi-public Spaces: e.g., Conference

4 I.BEHAVIOUR DETECTION WORKPLACE Interactions : foster idea flows leads to creation of social ties which impact loyalty could lead to collaboration Collaboration impacts productivity

5 I.BEHAVIOUR DETECTION WORKPLACE How can we capture and model employees face-to-face spontaneous encounters at workplace?

6 DETECTING FACE-TO-FACE ENCOUNTERS Imagery/Sound based detections

7 DETECTING FACE-TO-FACE ENCOUNTERS Imagery/Sound based detections Smartphones: 48% within arm reach 80% within 5 meters i-beacon

8 DETECTING FACE-TO-FACE ENCOUNTERS Imagery/Sound based detections Smartphones: 48% within arm reach 80% within 5 meters i-beacon Wi-Fi

9 FROM NOISE TO SOCIAL SIGNAL Radio Signal Detection WiFi is everywhere Using WiFi Probes (802.11) standard No additional infrastructure needed

10 FROM NOISE TO SOCIAL SIGNAL Network-Centric Using RSSI of the probes to cluster the nearby devices together. Radio Signal Detection Co-presence Detection Smart-phones as proxy for people Using WiFi Probes (802.11) standard No additional infrastructure needed

11 FROM NOISE TO SOCIAL SIGNAL Network-Centric Using RSSI of the probes to cluster the nearby devices together. Radio Signal Detection Co-presence Detection Encounter Inference Smart-phones as proxy for people Using WiFi Probes (802.11) standard No additional infrastructure needed Features based on human behaviour Using supervised machine learning models to detect encounters

12 DETECTING FACE-TO-FACE ENCOUNTERS

13 DETECTING FACE-TO-FACE ENCOUNTERS

14 DETECTING FACE-TO-FACE ENCOUNTERS

15 DETECTING FACE-TO-FACE ENCOUNTERS

16 Diary based in-the-wild deployment 8 Users 7 Hours 4 Access Points

17 Diary based in-the-wild deployment 8 Users 7 Hours 4 Access Points 86% precision 60% recall

18 TAKE AWAY MESSAGE Reduce burden on user Incorporate Sociologist theories

19 II.BEHAVIOUR DETECTION IN EVENTS

20 II.BEHAVIOUR DETECTION IN EVENTS n Space Analytics á People Analytics Spatio-Temporal Heat Map of the Venue Informed Space Management Personalised Space Recommendation Better Resource Management Crowd Flow Management Profiling Spatio-Temporal Behavior of Individual and Groups Extracting Communities and their Properties

21 CHALLENGES - Extremely crowded/noisy environment. Short lived event, - Large areas to monitor: infrastructure deployment could get costly using Bluetooth - Little control over crowds participation (e.g., having Bluetooth turned on). - Energy

22 SYSTEM Prototyped Badges - WiFi Badges emitting WiFi Management Frames (Probe) exploiting Sequence Control. - The badge composed of an ESP8266, a SoC consisting of an ARM-based CPU and a 2.4GHz Wi-Fi controller, a Freescale MMA8452Q accelerometer and 180mAH LiPo battery. - Algorithm to save energy by putting the badge to sleep based on accelerometer reading. - Badges can run 3 days without need to be recharged.

23 SYSTEM Prototyped Gateway - Detection of WiFi Badges through probes - Algorithm at the Gateway to capture specific radio signals. {mac, sequence, RSSI, timestamp} - No connectivity between the gateways or to internet. - Low cost

24 DEPLOYMENT - Web summit K participants sq meter - 30 Gateways, 9 zones. (One Gateway was taken). - Zones had functionalities: Booth area, Stages, and Social Lounges (limited access)

25 PARTICIPATION Entrepreneurs vs Investors - 85 Badges - Distributed between pre-selected entrepreneurs and investors.

26 SPATIAL ANALYSIS Collected Data - 7 Million probes from 290K unique MAC addresses. - Filtered 85% had less than 10 probes - 61 with valid observation: 34 investors, 27 entrepreneurs - 2.5K devices

27 SPATIAL ANALYSIS Collected Data - 7 Million probes from 290K unique MAC addresses. - Filtered 85% had less than 10 probes - 61 with valid observation: 34 investors, 27 entrepreneurs - 2.5K devices

28 SPATIAL ANALYSIS Proximity Detection - Ground truth data manually recorded - RSSI from each zone from 5 badges - Multi-class classification problem with k=9 non-overlapping zones - RSSI based feature vector of size 87 (mean, sd, frequency) - SVM with 80% training and 20% testing. F score= 0.92

29 PEOPLE ANALYTICS What impact spatial layout has on the opportunities available to the entrepreneurs to strengthen their social ties?"

30 PEOPLE ANALYTICS What impact spatial layout has on the opportunities available to the entrepreneurs to strengthen their social ties?"

31 PEOPLE ANALYTICS Investors Lounge 50% of Investors spent at least 60% of their time in Investors Lounge

32 PEOPLE ANALYTICS Investors Entrepreneurs

33 PEOPLE ANALYTICS Investors Entrepreneurs Investors took short strolls before returning to their lounge

34 TAKE AWAY MESSAGE Incorporate heterogeneity in design Create opportunities for encounter

35 THANK