A framework to model human behavior at large scale during natural disaster

These are the slide for the presentation I gave at IEEE MDM 2016 conference, Porto Portugal.

Transcript

1. A Framework To Model Human Behavior At Large Scale During Natural Disasters Jay Ghurye, Gautier Krings and Vanessa Frias-Martinez 1
2. Introduction and Problem Motivation • Millions of people get affected each year by natural disasters such as hurricanes, floods or tornadoes • Governments spend millions of dollars to mitigate the damages of disasters • Effective disaster management needs deep understanding of how humans react to disaster 2
3. Our Contribution We develop a framework: • To build behavioral and social network baseline models using CDR (Call Detail Record) data • Use these models to evaluate changes in human behavior after disaster • Evaluates real flood scenario in Rwanda using CDR data obtained from the major telecommunication carrier 3
4. Why CDR data? • Independent of handset used • Passive collection: need not be initiated by user as opposite to social network • Collect information at large scale as opposite to surveys and interviews 4
5. Framework 5
6. Data Pre-Processing 6
7. Information in CDR data • Each record is a tuple (caller, callee, location of tower, start time of call, end time, date) • Use location information to model mobility • Use caller – callee information to model ego social network 7
8. Data Preprocessing - Mobility Traces 8
9. Data Pre-Processing – Network Features 9 60% 40% 90% 10% Input and Output Degree Reciprocity Transitivity Use these to define friends User 1 User 2 F1 User 1 F2 User 1
10. Behavioral Baseline 10
11. Behavioral Baseline - Mobility 11 • Nth order Markov models ………… 1 2 3 4 K-2 K-1 K For N = 2 • Use Markov models of friends to predict the next location
12. Behavioral Baseline - Mobility Memoryless Baseline : MC(0) • Current state independent of previous states • Infers next location as the most frequent location visited in training data Time based memoryless baseline : TMC(0) • Considers the current state to be independent of previous states but dependent on day and time • Infers next location as the most frequent one visited at the given day of the week and time Compared our approach against two existing models in the literature 12
13. Behavioral Baseline - Mobility Exact location Top 30% location 13 Predicted Actual ? = Actual Top 30% predicted Belongs to? 1st 2nd 3rd 4th . . . . .
14. Disaster Analytics 14
15. Disaster Analytics - Displacements • Average weekly distance between predicted and observed location across all individuals • Changes in distribution of length of transitions lengths • Use Kolmogorov - Smirnov statistical test to evaluate behavioral differences 15
16. Disaster Analytics - Communications • Takes as input the baseline and post disaster social network features • Performs KS test to evaluate statically significant differences in features • Provides insight into how communications might change to reach out to others during a disaster 16
17. Region of Study 17
18. Dataset • CDR data of Rwanda during the period December 1st 2011 to June 30th 2012 1 • Total of 1.5 billion records for entire country • Anonymized to protect privacy of users • Considered the floods on 12th April 2012 in Musanze Province 18 1 Data similar to open data http://netmob.org
19. Evaluation – Behavioral Baselines 19 MODEL Exact Top 30% Friends with Top 30% MC(1) 40.21% 64.55% 66.44% MC(2) 44.32% 72.06% 73.68% MC(0) 44.47% 59.32% 60.40% TMC(0) 22.91% 29.88% 30.13%
20. Evaluation – Displacements 20
21. Evaluation - Transitions 21
22. Conclusion • Understanding human behavior is critical for disaster mitigation planning • Our framework helps to model user behavior during normal conditions and compare with behavior using disaster • Framework generates valuable information to improve emergency planning • Can be used for any type of disaster given if CDR data is available during the disaster 22
23. Thank You Jay Ghurye : jayg@cs.umd.edu Vanessa Frias-Martinez : vfrias@umd.edu 23