![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
mattbellHaving just moved into the Mission in San Francisco, I had an empty refrigerator but a hankering for food late at night. "Why not check out the best of the local taquerias?", I thought. The clear winner on Yelp was this place, with an average of 4.5 stars out of five with a mind-blowing 1100 reviews. I went there. The wait was horrendously long, and both the burrito and the soft taco I got (both recommended by reviewers) were horrible. The meat was far too dry and the burrito tasted soapy. It could have been an "off" day, but the place was crowded and others seemed to be eagerly eating their meals
Something can get five stars (or one star) for a lot of reasons. I'm guessing the 5-star reviewers like the place I tried for being open after the bars close and ridiculously cheap.
This is where collaborative filtering comes in. Collaborative filtering is a data analysis technique that analyzes everyone's reviews to find people similar to you. If I'm picky about food quality but not about ambience, it will find other people who are picky about food quality but not ambience. Then, these people's reviews can be weighted more heavily to provide more accurate ratings for me. Netflix uses this technique to provide personalized recommendations of movies.
I believe that implementing this technique would dramatically increase Yelp's usefulness to its users.
One nice thing for Yelp about collaborative filtering is that it encourages people to make a lot more reviews. If the system tells you that you will start getting personalized recommendations once you've reviewed 100 places, you're much more likely to start reviewing. Right now lots of people are freeloading off of Yelp by reading others' reviews but not posting their own. Adding collaborative filtering would dramatically change the incentive structure against freeloaders and get Yelp a pile of high quality user-generated content for free.
Another nice thing about collaborative filtering is that it would dramatically weaken the ability of restauranteurs to game the system by posting positive reviews on their own restaurants and negative reviews on their competitors'. The sham accounts used to make these reviews would carry a lot less weight because the accounts would likely not have enough reviews to influence collaboratively filtered ratings.
Something can get five stars (or one star) for a lot of reasons. I'm guessing the 5-star reviewers like the place I tried for being open after the bars close and ridiculously cheap.
This is where collaborative filtering comes in. Collaborative filtering is a data analysis technique that analyzes everyone's reviews to find people similar to you. If I'm picky about food quality but not about ambience, it will find other people who are picky about food quality but not ambience. Then, these people's reviews can be weighted more heavily to provide more accurate ratings for me. Netflix uses this technique to provide personalized recommendations of movies.
I believe that implementing this technique would dramatically increase Yelp's usefulness to its users.
One nice thing for Yelp about collaborative filtering is that it encourages people to make a lot more reviews. If the system tells you that you will start getting personalized recommendations once you've reviewed 100 places, you're much more likely to start reviewing. Right now lots of people are freeloading off of Yelp by reading others' reviews but not posting their own. Adding collaborative filtering would dramatically change the incentive structure against freeloaders and get Yelp a pile of high quality user-generated content for free.
Another nice thing about collaborative filtering is that it would dramatically weaken the ability of restauranteurs to game the system by posting positive reviews on their own restaurants and negative reviews on their competitors'. The sham accounts used to make these reviews would carry a lot less weight because the accounts would likely not have enough reviews to influence collaboratively filtered ratings.
![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
no subject
Date: 2010-03-11 12:19 am (UTC)no subject
Date: 2010-03-11 12:34 am (UTC)