Ads and the City
The document discusses using geographic distance and other factors to improve venue recommendation systems. It analyzes Foursquare data on venues and users in London to determine which categories of venues users tend to visit locations farther from home. It then develops a probabilistic model to calculate the likelihood a user will visit a venue based on their past likes, proximity, and visit frequency. The model is tested using naive Bayesian, Bayesian, and linear regression approaches. Results show geographic closeness strongly influences recommendations and incorporating domain knowledge significantly improves accuracy over basic collaborative filtering. The discussion notes the importance of understanding one's recommendation domain and how performance depends on venue category characteristics.
Ads and the City
- 1. Ads and the City: Considering Geographic Distance Goes a Long Way Diego Saez-Trumper1 Daniele Quercia 2 Jon Crowcroft 2 1 Universitat Pompeu Fabra, Barcelona 2 Computer Laboratory, University of Cambridge Dublin, September, 2012
- 3. Category #Venues #Users food 1,293 1,566 nightlife 1,075 1,207 travel 850 1,744 home/work/etc. 411 1,037 shops 362 878 arts&entertainment 348 841 parks&outdoors 184 363 education 49 117 Total 4,572 3,110 Table: London Foursquare Data
- 4. Given a venue, suggests guests
- 5. Context similar to target advertising (?) domain knowledge in people mobility
- 6. On people mobility (from the literature) distance matters likes might matter ¡°power users¡± are special p(go|like, close) ¡Ø pgo ¡¤ pclose ¡¤ plike
- 7. plike #venues visited by user u with rating lui p(like = lui |go) = total #venues visited by user u lui is ranking obtained from item-based CF algorithm.
- 8. pgo #venues visited by user u pgo = total #venues
- 9. pclose 1 pclose = k1 ¦Á dui
- 10. pclose 1 pclose = k1 ¦Á dui
- 11. pclose 1 pclose = k1 ¦Á dui Category ¦Á food 1.64 nightlife 1.61 travel (airports/trainstations) 2.22 home/work/etc. 1.62 shops 1.64 arts&entertainment 1.64 parks&outdoors 1.68 education 1.93 High ¦Á ¡ú travel farther
- 12. p(go|like, close) ¡Ø pgo ¡¤ pclose ¡¤ plike Naive Bayesian Bayesian Linear Regression
- 13. Results
- 14. Results
- 15. Results
- 16. Results 1.0 p_go p_close p_like 0.8 Naive Bayesian Linear Reg. 0.6 accuracy 0.4 0.2 0.0 Arts.and.Ent. Education Food HomeWork Nightlife Parks Shops Travel
- 18. When it does not work
- 19. When It Does not Work
- 20. When It Does not Work
- 21. Final Remarks results depend on venue category (different ¦Á and predictability) geographic closeness plays a very important role. domain knowledge signi?cantly improves recommendations results.
- 22. ¡°Understanding the speci?cs of your domain is critical to building a good recommender¡± Paul Lamere @ recsys¡¯12
- 23. Questions?