RALLY - rapid protest detection

Transcript

1. Tiananmen  Square  |  Instagram  @whothehellisalice Charlotte  Daniels RALLY RAPID PROTEST DETECTION
2. • How  do  you  break  a   story  first? • Leverage  social   media’s  crowd   sourced  information.
3. Data  from  the Instagram  API Feature  engineering -­‐ Captions -­‐ Hashtags Time  based   clustering Algorithm -­‐ Unsupervised   learning -­‐ Outlier  detection -­‐ OCSVM -­‐ nu  =  0.01 Classify  and  cluster Algorithm and data Return  classification, city,  date,  time,  map, &  highest  ranking  photo. Google  maps  API
4. Outlier detection Inliers Instagram  Data Inliers
5. Validation using news sources protests  detected 56/60
6. Validation using news sources
7. Who spots a protest first? Detection  Gap  Between  Media  and  RALLY RALLY  average  lead  time:   9.3  hours
9. THANK YOU
10. Photo references • https://instagram.com/p/3gcqsTAyma/ • https://instagram.com/p/2_TR5_syos/?taken-­‐ by=fernandochuyk • https://instagram.com/p/wZdwY5zcKs/?taken-­‐ by=messageframer • https://instagram.com/p/3zYr6bSrt9/ • https://instagram.com/p/3lAugqmjLC/ • https://scontent.cdninstagram.com/hphotos-­‐ xtp1/t51.2885-­‐ 15/e15/10817809_1515800758678909_445698922_n.jpg
11. • Positive  predictive  calculations • Probability  hour  tests  as  outlier  given  that  ground  truth   is  an  outlier • Probability  of  a  outlier • Probability  hour  tests  as  outlier  given  that  ground  truth   is  an  inlier • Probability  of  an  inlier | + = +| +| + +|& & +| +|& & = 0.93×0.005 0.93×0.005 + 0.0003×0.9945 | + = 0.93 10903 4 56 4 0  (inlier) 1 (outlier) 0  (inlier) 1 (outlier) Actual Test
12. Term frequency analysis