Disco

:fire: Collaborative filtering for Ruby

  • Supports user-based and item-based recommendations
  • Works with explicit and implicit feedback
  • Uses high-performance matrix factorization

Build Status

Installation

Add this line to your application’s Gemfile:

gem 'disco'

Getting Started

Create a recommender

recommender = Disco::Recommender.new

If users rate items directly, this is known as explicit feedback. Fit the recommender with:

recommender.fit([
  {user_id: 1, item_id: 1, rating: 5},
  {user_id: 2, item_id: 1, rating: 3}
])

IDs can be integers, strings, or any other data type

If users don’t rate items directly (for instance, they’re purchasing items or reading posts), this is known as implicit feedback. Leave out the rating, or use a value like number of purchases, number of page views, or time spent on page:

recommender.fit([
  {user_id: 1, item_id: 1, value: 1},
  {user_id: 2, item_id: 1, value: 1}
])

Use value instead of rating for implicit feedback

Get user-based (user-item) recommendations - “users like you also liked”

recommender.user_recs(user_id)

Get item-based (item-item) recommendations - “users who liked this item also liked”

recommender.item_recs(item_id)

Use the count option to specify the number of recommendations (default is 5)

recommender.user_recs(user_id, count: 3)

Get predicted ratings for specific items

recommender.user_recs(user_id, item_ids: [1, 2, 3])

Get similar users

recommender.similar_users(user_id)

Examples

MovieLens

Load the data

data = Disco.load_movielens

Create a recommender and get similar movies

recommender = Disco::Recommender.new(factors: 20)
recommender.fit(data)
recommender.item_recs("Star Wars (1977)")

Ahoy

Ahoy is a great source for implicit feedback

views = Ahoy::Event.
  where(name: "Viewed post").
  group(:user_id, "properties->>'post_id'"). # postgres syntax
  count

data =
  views.map do |(user_id, post_id), count|
    {
      user_id: user_id,
      item_id: post_id,
      value: count
    }
  end

Create a recommender and get recommended posts for a user

recommender = Disco::Recommender.new
recommender.fit(data)
recommender.user_recs(current_user.id)

Storing Recommendations

Disco makes it easy to store recommendations in Rails.

rails generate disco:recommendation
rails db:migrate

For user-based recommendations, use:

class User < ApplicationRecord
  has_recommended :products
end

Change :products to match the model you’re recommending

Save recommendations

User.find_each do |user|
  recs = recommender.user_recs(user.id)
  user.update_recommended_products(recs)
end

Get recommendations

user.recommended_products

For item-based recommendations, use:

class Product < ApplicationRecord
  has_recommended :products
end

Specify multiple types of recommendations for a model with:

class User < ApplicationRecord
  has_recommended :products
  has_recommended :products_v2, class_name: "Product"
end

And use the appropriate methods:

user.update_recommended_products_v2(recs)
user.recommended_products_v2

For Rails < 6, speed up inserts by adding activerecord-import to your app.

Storing Recommenders

If you’d prefer to perform recommendations on-the-fly, store the recommender

bin = Marshal.dump(recommender)
File.binwrite("recommender.bin", bin)

You can save it to a file, database, or any other storage system

Load a recommender

bin = File.binread("recommender.bin")
recommender = Marshal.load(bin)

Algorithms

Disco uses high-performance matrix factorization.

Specify the number of factors and epochs

Disco::Recommender.new(factors: 8, epochs: 20)

If recommendations look off, trying changing factors. The default is 8, but 3 could be good for some applications and 300 good for others.

Validation

Pass a validation set with:

recommender.fit(data, validation_set: validation_set)

Cold Start

Collaborative filtering suffers from the cold start problem. It’s unable to make good recommendations without data on a user or item, which is problematic for new users and items.

recommender.user_recs(new_user_id) # returns empty array

There are a number of ways to deal with this, but here are some common ones:

  • For user-based recommendations, show new users the most popular items.
  • For item-based recommendations, make content-based recommendations with a gem like tf-idf-similarity.

Data

Data can be an array of hashes

[{user_id: 1, item_id: 1, rating: 5}, {user_id: 2, item_id: 1, rating: 3}]

Or a Rover data frame

Rover.read_csv("ratings.csv")

Or a Daru data frame

Daru::DataFrame.from_csv("ratings.csv")

Faster Similarity

If you have a large number of users/items, you can use an approximate nearest neighbors library like NGT to speed up item-based recommendations and similar users.

Add this line to your application’s Gemfile:

gem 'ngt', '>= 0.3.0'

Speed up item-based recommendations with:

model.optimize_item_recs

Speed up similar users with:

model.optimize_similar_users

This should be called after fitting or loading the model.

Reference

Get the global mean

recommender.global_mean

Get the factors

recommender.user_factors
recommender.item_factors

Credits

Thanks to:

  • LIBMF for providing high performance matrix factorization
  • Implicit for serving as an initial reference for user and item similarity
  • @dasch for the gem name

History

View the changelog

Contributing

Everyone is encouraged to help improve this project. Here are a few ways you can help: