What would a local tell you to do?

We love to feel like a local when we travel. And we believe that with Triposo on your phone you can feel like a local. You can wander around without the fear of being lost because our offline maps will always get you home, you have suggestions that will tell you about the best places in the area and we make sure you will know about local events happening.

During the Scotland Jamboree I set out to find out what factors determine if a local is likely to suggest a restaurant or a bar to a travelers.

The methodology I used was as follows.

  • I started with a training set. I selected a number of points of interest in London that locals were very unlikely to suggest (tourist traps) and a list of places locals were likely to suggest. The lists were based on personal experience as well as research.
  • I then collected a number of statistics about each of the places. I decided to look at some new statistics (like number of Tripadvisor reviews or number of tips on Foursquare) we don’t use in our current scoring - and add scoring as one of them.

The statistics I used were:

  • Number of tips added on Foursquare
  • Number of likes on Facebook
  • Number of talking about on Facebook
  • Number of checkins on Facebook
  • Number of Tripadvisor reviews
  • Number of negative mentions for a specific set of dishes and tags in a set of 5M reviews
  • Number of positive mentions for a specific set of dishes and tags in a set of 5M reviews
  • Number of geotagged images taken in the area from a database of 100 M pictures
  • Price level indication
  • Triposo score
  • Is it part of a chain (ie do we have a significant number of places in our db with the same name).

I then created formula with these factors. By minimizing the sum of squared differences I could optimize the algorithm to predict the scores in the training set.

The following graph shows how different factors impact how likely locals are to recommend a place:

Algorithm in action


San Francisco


New York


A few notes:

* Facebook statistics seemed to be too noisy to use, mostly because the activity of the page owner has a stronger influence. In a next round I would like to compare Facebook popularity with Twitter popularity to see if a difference exists between owners that promote their place on Facebook rather than on Twitter and if this correlates with certain types of places.

* I ended up adding some corrections to the algorithm because my initial set did not have any real outliers: they all had perfect data. Which means they all had scores for all of the factors above. When you then use the algorithm on 500K points of interest you see that missing matches influences scores to much.

* The training set was probably too small for scientific purposes or even for producing a algorithm we can use tomorrow in our apps. But all in all this shows real promise. It does bring up really good, slightly untouristy places in all the destinations we tried!