beer-reco

It is always fun to try new beers and we look for recommendations when are out to try a new variety. This tiny project is an attempt to use some reviews (~1.5 Million) of various microbrews across the world collected from www.beeradvocate.com and train Machine Learning models to give recommendations based on alcohol value, aroma/taste/body etc., Python Data Science Stack and Machine Learning algorithms as implemented in Scikit-learn are used. Also answers the questions in this blog: http://tcbanalytics.com/blog/testing-data-skills.html#.WFWApHeZOb-

Collection of Notebooks

beer-reco repository is a collection of jupyter notebooks (.ipynb) that explain the flow of thought and answers few questions in the blog above by ranking beers based on attributes and to generate some recommendations.

1. Sorting and Ranking.ipynb

Data Import, basic sorting and ranking of data in Pandas DataFrames and visualization. Example:

• which brewery produces strongest beers by ABV%?

2. General Recommendations.ipynb

Based on the overall reviews from reviewers across the world, another simple strategy is to rank the beers based on overall reviews. This requires a bit more work than simple sorting because of inhomogenieties in the data. All beers are not rated by all reviewers and ratings are distributed around a certain mean value with some strandard deviation. After the data cleaning step to retain statistically significant values (to predict with 95% confidence interval), reviews can be sorted and ranked to provide general recommendations Examples:

• Beer or top five beers with best overall review
• Top five beers with best aroma.. etc

3.Personalized Recommendations.ipynb

Personalized recommendations using collaborative filtering. Besides general recommendations, this notebook trains a statistical model using low rank matrix factorization and collaborative filtering. User preferences are collected and the model provides personalized recommendations.

4. Beer Rating Prediction.ipynb

Regression models to predict overall rating of a beer based on sample means of aroma, appearance, taste and palate. Includes cross-validation by splitting training and test data sets using KFold split and averaging the model parameters over all sets. Includes some utility functions and scikit-learn method wrapper class.

5. Beer Style Classification.ipynb (Coming Soon)