Recommendable is a gem that allows you to quickly add a recommendation engine for Likes and Dislikes to your Ruby application using my version of Jaccardian similarity and memory-based collaborative filtering.
Bundling one of the queueing systems above is highly recommended to avoid having to manually refresh users' recommendations. If you bundle Sidekiq, you should also include 'sidekiq-middleware' in your Gemfile to ensure that a user will not get enqueued more than once at a time. If bundling Resque, you should include 'resque-loner' for this. As far as I know, there is no current way to avoid duplicate jobs in DelayedJob. Queueing for Torquebox is also supported.
Add the following to your application's
Please note that you currently must need to place Recommendable below your ORM and queueing system in the Gemfile. If you are using Sidekiq and ActiveRecord, please place
gem recommendable below both
gem 'rails' and
After bundling, you should configure Recommendable. Do this somewhere after you've required it, but before it's actually used. For example, Rails users would create an initializer (
require 'redis' Recommendable.configure do |config| # Recommendable's connection to Redis. # # Default: localhost:6379/0 config.redis = Redis.new(:host => 'localhost', :port => 6379, :db => 0) # A prefix for all keys Recommendable uses. # # Default: recommendable config.redis_namespace = :recommendable # Whether or not to automatically enqueue users to have their recommendations # refreshed after they like/dislike an item. # # Default: true config.auto_enqueue = true # The number of nearest neighbors (k-NN) to check when updating # recommendations for a user. Set to `nil` if you want to check all # neighbors as opposed to a subset of the nearest ones. Set this to a lower # number to improve Redis memory usage. # # Default: nil config.nearest_neighbors = nil # Like kNN, but also uses some number of most dissimilar users when # updating recommendations for a user. Because, hey, disagreements are # just as important as agreements, right? If `nearest_neighbors` is set to # `nil`, this configuration is ignored. Set this to a lower number # to improve Redis memory usage. # # Default: nil config.furthest_neighbors = nil # The number of recommendations to store per user. Set this to a lower # number to improve Redis memory usage. # # Default: 100 config.recommendations_to_store = 100 end
The values listed above are the defaults. I recommend playing around with the
nearest_neighbors setting. A higher value will provide more accurate recommendations at the cost of more time spent generating them.
If your application uses multiple ORMs, you must configure Recommendable to use the correct one. For example:
Recommendable.configure do |config| config.orm = :active_record end
Important: in case of
active_record with id of type
In your model that will be receiving recommendations:
class User recommends :movies, :books, :minerals, :other_things # ... end
To ensure that users' recommendations are processed after they rate items, make sure your bundled queue system is running:
# sidekiq $ [bundle exec] sidekiq -q recommendable # resque $ QUEUE=recommendable [bundle exec] rake environment resque:work # delayed_job $ [bundle exec] rake jobs:work
That's it! Please note, however, that currently only one model may receive recommendations.
Recommendable requires Redis to deliver recommendations. The collaborative filtering logic is based almost entirely on set math, and Redis is blazing fast for this.
NOTE: Your Redis database MUST be persistent. All ratings are stored permanently in Redis. If you're worried about Redis losing data, keep backups.
For Mac OS X users, homebrew is by far the easiest way to install Redis. Make sure to read the caveats after installation!
$ brew install redis
For Linux users, there is a package on apt-get.
$ sudo apt-get install redis-server $ redis-server
Redis will now be running on localhost:6379. After a second, you can hit
ctrl-\ to detach and keep Redis running in the background.
I'll let Randall Munroe of XKCD take this one for me: