safeparking

Transcript

1. SafeParking   Jiayi  Xie   Insight  Data  Science  2016  
2. Finding  safe  parking  can  be  challenging     In  Chicago,  ~24,000  car  related  crimes  per  year    
3. Finding  safe  parking  can  be  challenging     In  Chicago,  ~24,000  car  related  crimes  per  year     Half  of  the  crime  à  top  20%  riskiest  locaCons      
4. Historical  Vehicle  Crimes     Kernel  Density  Es>ma>on   (KDE)   +    2  Point  Correla>on   Risk  Map    
5. d   Length  scale  ~  250m  (a  few  blocks)   è  Crime  tend  to  happen  closer  to   each  other  within  a  few  blocks!   Two  Point  Correla>on  Analysis    
6. Valida>on     Compare  risk  scores  computed  from  odd  and  even  days  of   the  year  
7. Jiayi  Xie   PhD  in  Geophysics  
9. Do  crimes  tend  to  happen  at  certain  >me?       Yes!     Most  of  them  happen  when  you  are   having  fun   Periodicity  of   Crimes  (from   FFT):   •  0.5  day   •  1  day   •  7  days  
12. d   Two  Point  Correla>on  Analysis     ij number of pairs in D and R is n ( n 1) / 2 and r ( r the clustering at the angular scale ✓ , count the num separations in the interval ✓ ± ✓/ 2. Let DD and RR and R , respectively. The ‘natural’ estimator of w ( ✓ ) i w1 = DD n ( n 1) / 2 RR r ( r 1) / 2 RR r ( r 1) / 2 = r ( r n ( n To reduce the e↵ect of statistical fluctuations in RR (say, 10) random fields, and use the average number h As shown by Landy & Szalay [2], w1 is biased (i.e. even when there is no clustering). A much better the cross-correlation statistic DR , that is the numbe ✓ ± ✓/ 2, with one point taken from D and the othe The Landy–Szalay estimator is then computed as2 DD 2 DR RR r ( r 1) / 2 To reduce the e↵ect of statistical fluctuations (say, 10) random fields, and use the average nu As shown by Landy & Szalay [2], w1 is bias even when there is no clustering). A much the cross-correlation statistic DR , that is the ✓ ± ✓/ 2, with one point taken from D and t The Landy–Szalay estimator is then computed w3 = DD n ( n 1) / 2 2 DR nr + RR r ( r 1) / 2 RR r ( r 1) / 2 = It is advisable to replace RR and DR by the m In the actual calculations, DD , RR , and D counting the number of pairs in m bins of number of pairs with 0 < ✓  ✓ , DD (2) the on, up to DD ( n ), which is the number of pair
13. d   Two  Point  Correla>on  Analysis     Random/uniform   Non-­‐uniform