Decider

Yet Another Ruby Machine Learning Library

Manifesto

There are other ruby machine learning libraries out there:

• github.com/cardmagic/classifier
• github.com/pauldix/basset
• More?

So why another one?

• You can install it and try it in irb right away. You don’t need to learn how a half-dozen classes work to get started:

        c = Decider.classifier(:spam, :ham)
        c.spam << "some spammy text"
        c.ham << "some hammy goodness"
        c.spam?("more spammy text")
        # => true
  

The default configuration is about 96% accurate as an email spam classifier.

• You can control how it processes its input easily. Decider has built-in support for plain text and URIs, stemming words, stop word removal, and n-grams. All of these can easily be combined at your option (see "Getting Started" below for a quick example). Additional tokenization strategies or support for non-text document types can be added with a minimum of hassle.
• Persist (Save) with Moneta. Pretty much any storage mechanism that’s available in ruby is supported. Save to a database and implement distributed classification if you like.
• Clustering Analysis. Useful for recommendation algorithms. (In Progress)

Getting Started

  c = Decider.classifier(:spam, :ham) do |doc|
    doc.plain_text
    doc.ngrams(2..3)
    doc.stem
  end
  c.spam << "buy viagra, jerk" << "get enormous hot dog for make women happy"
  c.ham << "check out my code on github homie" << "let's go out for beers after work"
  p c.spam?("viagra for huge hot dog")
  # => true
  puts "term frequencies:"
  puts "spam: #{c.spam.term_frequency.inspect}"
  puts "ham:  #{c.ham.term_frequency.inspect}"
  puts ""
  p c.scores("let's write code and drink some beers")
  # => {:spam=>0.0, :ham=>1.0}
  p c.classify("let's write code and drink some beers")
  # => :ham

Performance

Decider has several benchmarks that also double as integration tests. These are run regularly and used to pinpoint CPU and RAM bottlenecks.

Decider does a lot of math and is fairly computationally intensive, so you want all the extra speed you can get. It is regularly tested with Ruby 1.9 and Jruby. I highly recommend using one of these Ruby implementations if at all possible if you plan on doing anything serious with Decider.

Also keep in mind that your dataset should reside entirely in memory or else you’ll hit a brick wall.