Eric Horvitz Microsoft Research. PerCom 2011 Seattle, WA

Transcription

1 Eric Horvitz Microsoft Research PerCom 2011 Seattle, WA

2 What might we learn about human behavior and of world phenomena from logs of the activities of people?

3 How can we harness data about people while acting in accordance with preferences about privacy?

4 How might we leverage social connectivity and awareness in the crowdsourcing of solutions?

5 Data from widely used communication systems GPS, Wifi, cell tower, IM, searching, app usage Insights about people Sensing phenomena in world

6 Buddy Conversation Leskovec & Horvitz, 2008

7 Study: One month of Messenger data 245 million users logged in 180 million users engaged in conversations Communication graph (two-way) > 30 billion conversations > 255 billion messages exchanged 1.3 billion edges 4.5 terabytes All data anonymized. No message text Leskovec & Horvitz, 2008

8 For each conversation between geo points (A,B) we increase the intensity on the line between Leskovec A and & Horvitz, B 2008

9 Width: Numbers of conversations Leskovec & Horvitz,

10 Each point represents number of users at location 10

11 Users per capita Leskovec & Horvitz,

12 Number of steps between pairs of people Avg. path length % of others can be reached in < 8 hops Leskovec & Horvitz, 2008

13 Conversations per acquaintance Longer links used more Conversations per acquaintance with distance Leskovec & Horvitz, 2008

14 User self-reported age Young people communicate with same age High Low 14

15 User self-reported age Older people exchange more messages per session. High Low 15

16 User self-reported age Young people converse more quickly High Low 16

17

18

19 Feb. 3,

20 3 years of all cell communications in Rwanda 140 cell towers, 6 days: 10,527,799 calls Kapoor, Eagle, Horvitz, 2010

21 Normal trend at each cell tower p a i t No Event)~N(m i, σ i 2 ) estimate from history

22 Modeling deviations from the trend p( ai Event ) ~ N( m i (1 i ), i ) i 2 ( ex xi ) ( ey yi ) 2 x i,y i,,,e x,, e y epicenter

23 Error: km

24 Opportunities for Assistance (Day 2)

25 Ideal Reconnaissance (Day 2)

26

27

28 Case library ~1,000,000 km ~100,000 trips

29 maps.bing.com * m.bing.com 72 cities across North America Flows assigned to ~60 million streets every few minutes

30 maps.bing.com * m.bing.com

31

32 Data: AM/PM commutes to & from Microsoft Kamar & Horvitz, 2009

33 Commute request Rideshare queued Rideshare starts Single rider starts

34 Normal commute Computed rideshares

35 System Efficiency "What If?" Studies Efficiency on number of commutes Efficiency on total cost Number of agents Number of participants Kamar & Horvitz, 2009

36 Challenge: Understanding acceptance, perceptions, social considerations Address concerns, leverage opportunities Trusted organizations Referral, reputation - e.g., existing online social networks (e.g., link distance bounds)? Kamar & Horvitz, 2009

37

38 How can we harness data about people while acting in accordance with preferences about privacy?

39 Sensing guided by preferences & contracts Cost-benefit analysis, k-anonymity, differential privacy, sampling with risk Protected sensing & personalization Local sensing and execution, use population data when available

40 We can share a little bit & have a great system for all. Road demand + flow uncertainty used to compute value of sensing Availability Preferences Demand Phenomenon Krause, Horvitz, Kansal, F. Zhao, 2008

41 Compute value of information for selective sensing. s 4 s 1 s 2 s 3 s 6 s 5 s 7 s 8 Predict unmonitored flows from sensed flows e.g., p(s 5 S 1, S 9 ) s 11 s 9 s 9 s 10 s 12 Krause, Horvitz, Kansal, F. Zhao, 2008

42 Sensing under privacy budget can perform well. Demand-weighted variance Random road cells cells cells Optimize road Number of observations Low selection noise even with 53 cells

43 Tradeoffs: Balancing Utility & Sensitivity Balance and Preferences Value of information Sensitivity of private information Identifying a sweet spot? Maximize value at minimum sensitivity Krause & Horvitz, 2010

44 Freq Freq Example: Web search Value of personal data in enhanced service: Demographic data Search history (same query, searches per day?) Topic interests (ever visited business, kids, etc. website) User behavior (location, ever changed zip)? Query: sports Pages Entropy H = 2.6 Country USA H = 1.7 Entropy Reduction: 0.9 Krause & Horvitz, 2010

45 Reduction in click entropy Privacy cost (p log(1-p)) Utility - Cost Trading off value & cost Value: Diminishing returns (submodular) Cost: Accelerating (supermodular) Optimization 1.4 Greedy forward selection for utility (Lazy) Greedy forward selection λ = Optimal tradeoff! 0 no occ adult age whour gdr home wday ctry kids refs bus comp world arts reg 0 wday whour reg bus adult world ctry gdr arts comp refs kids home age occ all -0.2 none occ home whour age wday gender reg bus world adult arts country comp ref kids More observations More observations More observations Krause & Horvitz, 2010

46 Reduction in click entropy Diminishing Returns on Utility 1.4 Greedy forward selection for utility no occ adult age whour gdr home wday ctry kids refs bus comp world arts reg Krause & Horvitz, 2010

47 Utility - Cost Finding the Sweet Spot 0.8 (Lazy) Greedy forward selection Optimal tradeoff! -0.2 none occ home whour age wday gender reg bus world adult arts country comp ref kids Krause & Horvitz, 2010