We all love beers. And today, we face an unprecedented variety of options. We are overwhelmed by too many choices and couldn’t decide: What should I drink next?
In our Hack Reactor thesis project, I build a machine learning server using PredictionIO as a recommendation engine for beer. Broadly speaking, the app is based on the following two strategies:
- Content filtering approach — using each beer to characterize its nature. For example, we use breweryDB to find out the style of a specific beer, alcohol by volume (abv) and the International Bittering Units (ibu). Then we query the database to find beers with a similar style.
- Collaborative filtering approach — replies only on past user behaviour, i.e. your beer ratings. It is based on matrix factorization techniques with alternating least squares (ALS) algorithm. We characterize both beers and users by vectors of factors inferred from beer rating patterns. High correspondence between beer and user factors leads to a recommendation.
The advantage of using matrix factorization is that it allows incorporation of additional information. When explicit feedback is not available (i.e. your ratings), we can infer user preferences using implicit feedback, such as your browsing history and search pattern.
As a result, OnTapp matches you with beers that suit your taste. To get a recommendation of beer and try our demo, please visit: http://ontappapp.com/
Originally published at victorleungtw.com on January 20, 2015.