insightweek5

from sunday, June 28

Transcript

1. Parental Guidance: finding Yelp reviews written by parents -Amanda Ford The opportunity: •  People value reviews written by others like them, esp. parents •  Can you identify parents with info the user has already provided? •  Can you build a useful tool exclusively with input from a given cohort? 1   Amanda  Ford  
2. •  Yelp’s Data Challenge: 1.5+ million reviews from 300,000+ users across 6 U.S. cities (~2Gb) •  How to identify parents? 2   Words: ‘kids’, as in “my wife and kids” ‘baby’, as in “our baby loved it!” Categories: ‘Pediatricians’: “Dr. L is the best!” ‘Preschools’: “We love St. Marks!” Signal Words: ‘kids’, as in “way too many loud kids” ‘baby’, as in “Oh baby that was hot!” Categories (cross-listed) ‘General Practice’: “Dr. L is the best urologist’ ‘Religious Organizations’: “We love St. Marks Sunday worship services ” Noise Amanda  Ford   The Data
3. •  Supervised learning, k-Means clustering, word frequency analysis all suffered from: –  Noise –  Lack of clear class of non-parents •  Simple bigram formula worked best –  Searched for possessive word(s) + kid word –  “My kid”, “Our son”, “My wife and baby” •  Once a user was tagged as a parent on one review, that tag was propagated across all their other reviews 3   Amanda  Ford   Algorithm
4. •  > 20% of Yelp reviews written by parents •  Parents rate differently 4   Amanda  Ford   Data Story
5. 5   Amanda  Ford   Parent Ratings
6. 6   Amanda Ford
7. BACKUPS   7   Amanda  Ford  
8. Unsupervised  Learning?   Plot  frequency  of  broader  range  of  kid  words  (“my  kids”  AND  “kids”,  “babies”,  “stroller”  etc)   vs.  how  oRen  someone  reviews  a  business  in  a  “kid  category”.     1.  ALL  users:  BLUE   2.  Maybe  parents  (said  “my  kid”  or  similar,  OR  reviewed  a  kid  business  such  as  a  daycare)   3.  Definitely  parents  (said  “my  kid”  or  similar)   Look  for  clusters!!!       8   Amanda  Ford  
9. Search  for  Dis[nguishing  Bigrams   •  Bigrams  will  biggest  difference  between  parents  and   other  aren’t  obviously  parent  related     •  Parents  less  likely  to  say:   –  ‘have  been’   –  ‘when  I’   –  ‘they  were’   –  ‘have  to’   •  Parents  more  likely  to  say:   –  ‘out  of’   –  ‘a  bit’   –  ‘very  good’   –  ‘that  I’   Amanda  Ford   9  
10. 10   (More to come, searching for *which* category/ies have the greatest delta between parents and overall set) Tool is Useful: Parents can review differently than non-parents DRAFT   Amanda  Ford  
11. Signal   Noise   Words:    ‘kids’,  as  in  “my  wife  and  kids”    ‘baby’,  as  in  “our  baby  loved  it!”     Categories:    ‘Pediatricians’:  “Dr.  L  is  the  best!”    ‘Preschools’:  “We  love  St.  Marks!”    ‘Kid’s  ac[vi[es’:  “The  pedng  zoo    was  great!”     Words:    ‘kids’,  as  in  “way  too  many  loud  kids”    ‘baby’,  as  in  “Oh  baby  that  soup  was  hot!”     Categories:    ‘General  Prac[ce’:  “Dr.  L  is  the  best  urologist’    ‘Religious  Organiza[ons’:  “We  love  St.    Marks  Sunday  worship  services  ”    ‘Kid’s  Ac[vi[es’:    “Why  you’d  take  a  girl  to    a  pedng  zoo  on  a  first  date  is  beyond  me”      Also  employees?  former  users?   11   Amanda  Ford  
12. Methods:   •  You  have  a  set  of  people  who  are  definitely  parents   (using  selec[on  for  obvious  parent  phrases:  “my   son”,  “my  daughter”,  “my  kid”,  “my  children”)   •  You  could  just  select  above,  which  captures  ~10%  of   reviews     –  (underselec[on  beker  than  overselec[on)   –  Poten[ally  misses  a  large  group  of  parent  users     •  How  to  extend??     –  SUPERVISED  learning:  Could  use  that  to  train  a  model  to   look  for  more  parents   •  You  know  who’s  definitely  a  parent,  but  you  don’t  know  who’s   NOT  a  parent   •  (Not  so  viable)     12   Amanda  Ford  
13. NEXT  Step:  Find  more  words!   •  Build  a  larger  set  of  words  that  uniquely   iden[fy  parents,  e.g.  “diaper”,  “breast-­‐ feeding”,  “Baby  Bjorn”,  etc   •  Find  those  words  by  doing  frequency  analysis   of  users  I  already  know  are  parents  vs.   everyone  else   13   Amanda  Ford