This document describes a city attractions recommender system called TokyoGo that was developed using data collected from Foursquare APIs and web scraping. The system used machine learning techniques like NMF and DBScan clustering to analyze over 263,000 venue records and user tips to generate topic tags and recommendations. Evaluation of the model found it was able to distinguish preferences of local and foreign visitors to some extent. Further improvements and applications of the system are discussed.

1. TokyoGo
City Attractions Recommender
Wan-Ru Yang
Dec, 2016

2. Introduction Data Coll ResultsAnalysis
263103 Records
Discussion

3. Foursquare
API Query
Webscraping
AWS S3
Photos
PostgreSQLVenues Mongodb
Users
Tips
Introduction Data Coll ResultsAnalysis
Image Tagging
TensorFlow/ Spark EMR

4. Feature Extraction
Tipgs
DBScan
NMF
Cluster Model User Interface
Web App
Location
Venue Stats
Introduction Data Coll ResultsAnalysis Discussion

5. Foursquare
API Query
Webscraping
AWS S3
Photos
PostgreSQL
Mongodb
Introduction Data Coll ResultsAnalysis Discussion
Tipgs
DBScan
NMF
Recommender
Web App
LocationVenues
Tips
Users
Data Storage Features Process

6. Introduction Data Coll ResultsAnalysis Discussion

7. Introduction Data Coll ResultsAnalysis Discussion
Topic tag Top words
theme tour
Famous shrine, totoro,
national park …
Game or outdoor Video game, sunshine, bandit
History Garden, manju, oldtokyo
Culture Shinkansen, coast, market
theme park and shopping
Indianajons, waterfall,
waterpark

8. Recommended VenuesUser Input
Introduction Data Coll ResultsAnalysis Discussion

9. Introduction Data Coll ResultsAnalysis Discussion

10. Next steps:
• Improve the model with user – user similarity
• Include the seasonality and full tips data
• Train a neural network model to tag the image content
Introduction Data Coll ResultsAnalysis Discussion
• The method applied was able to distinguish (to a certain extent)
preferences of different groups (local, visitors from other areas in Japan,
and forigner travelers).
• My recommender system product of this project will include only the top
200 venues of each visitor source group (sum up to ~ 500 venues) as an
toy example that can be deployed on a small amazon instance. The
framework can be extended when more data available, and he business
features and A/B testing evaluation can be added.
• The NMF analysis indicates visitors to all the venues tend to mention
some food, which also indicates that food is an important element that
shared among all city attractions! Restaurant recommender is not the
topic of this project, but I am expecting to see interesting patterns among
different tourist sources in Tokyo.

11. https://github.com/WanRuYang
https://zuya.siraya.net
https://www.linkedin.com/in/WanRuYang