Bunch of modifications

Rakefile

@@ -43,7 +43,7 @@ end
 
 task :default => :test
 
-require 'rake/rdoctask'
+require 'rdoc/task'
 Rake::RDocTask.new do |rdoc|
   if File.exist?('VERSION')
     version = File.read('VERSION')

lib/k_means/centroid.rb

@@ -30,13 +30,34 @@ def create_ranges(nodes, dimensions)
       ranges
     end
   end
-  
-  attr_accessor :position
-  
+
+  attr_accessor :position, :nodes
+
   def initialize(position)
     @position = position
+    @mean_distance = nil
+    @nodes = []
   end
-
+
+  def mean_node_distance
+
+    return @mean_distance if @mean_distance
+
+    total_dist = 0.0
+    total_nodes = @nodes.size
+
+    total_dist = @nodes.reduce(0){|sum, node| sum + node.best_distance}
+
+    if total_nodes > 0
+      @mean_distance = total_dist/total_nodes
+    else
+      # if there no nodes in cluster, so the centroid is bad
+      @mean_distance = 1.0/0.0
+    end
+
+    @mean_distance
+  end
+
   # Finds the average distance of all the nodes assigned to
   # the centroid and then moves the centroid to that position
   def reposition(nodes, centroids)

lib/k_means/k_means.rb

@@ -1,17 +1,53 @@
 require 'ext/object'
+require 'pry'
 
 class KMeans
 
-  attr_reader :centroids, :nodes
+  attr_reader :centroids, :nodes, :max_iterations, :max_tries
 
   def initialize(data, options={})
-    distance_measure = options[:distance_measure] || :euclidean_distance
-    @nodes = Node.create_nodes(data, distance_measure)
-    @centroids = options[:custom_centroids] ||
-      Centroid.create_centroids(options[:centroids] || 4, @nodes)
+
+    @max_tries = options[:max_tries] || 10
+
+    @best_DBI = Float::INFINITY
+    @best_centroids = nil
+    @distance_measure = options[:distance_measure] || :euclidean_distance
+    @nodes = Node.create_nodes(data, @distance_measure)
+    @max_iterations = options[:max_iterations] || 100
     @verbose = options[:verbose]
 
-    perform_cluster_process
+    @centroid_count = options[:centroids] || 4
+
+    raise "Too many centroids(#{@centroid_count}) for #{@nodes.size} nodes" if @centroid_count > @nodes.size
+
+    @max_tries.times do |n|
+      reset_nodes!
+      @centroids = options[:custom_centroids] ||
+        Centroid.create_centroids(@centroid_count, @nodes)
+
+      perform_cluster_process
+      dbi = davies_bouldin_index
+      if dbi < @best_DBI
+        p "#{n} #{dbi.round(2)} vs #{@best_DBI.round(2)}" if options[:debug]
+        @best_DBI = dbi
+        @best_centroids = @centroids
+      end
+    end
+
+
+    if @best_centroids != nil
+      # can not fit nodes to all clusters
+      @centroids = @best_centroids
+    end
+
+
+
+    update_nodes(true) # force update
+    assign_nodes_to_centriods
+    place_nodes_into_pockets
+
+    #binding.pry if options[:debug]
+
   end
 
   def inspect
@@ -22,18 +58,42 @@ def view
     @centroid_pockets
   end
 
+  # Davies–Bouldin index: http://en.wikipedia.org/wiki/Cluster_analysis#Evaluation_of_clustering
+  def davies_bouldin_index(centroids = @centroids)
+    c_sz = centroids.size
+    db_index = 0
+    (0..(c_sz - 1)).each do |i|
+      max_db_index = -1.0/0
+      (0..(c_sz - 1)).each do |j|
+        if i != j
+          centroid_dist = centroids[i].position.send(@distance_measure, centroids[j].position)
+          sum_mean_nodes = centroids[i].mean_node_distance + centroids[j].mean_node_distance
+          max_db_index = [max_db_index, sum_mean_nodes / centroid_dist].max
+        end
+      end
+      db_index += max_db_index
+    end
+    return db_index/c_sz
+  end
+
+  def reset_nodes!
+    if @nodes
+      @nodes.each{|n| n.reset!}
+    end
+  end
+
   private
 
   def perform_cluster_process
     iterations, updates = 0, 1
-    while updates > 0 && iterations < 100
+    while updates > 0 && iterations < max_iterations
       iterations += 1
       verbose_message("Iteration #{iterations}")
       updates = 0
       updates += update_nodes
       reposition_centroids
     end
-    place_nodes_into_pockets
+    assign_nodes_to_centriods
   end
 
   # This creates an array of arrays
@@ -51,10 +111,17 @@ def place_nodes_into_pockets
     @centroid_pockets = centroid_pockets
   end
 
-  def update_nodes
+  def assign_nodes_to_centriods
+    @nodes.each_with_index do |node, node_index|
+      #Store closest nodes in the centroid object
+      node.closest_centroid.nodes << node
+    end
+  end
+
+  def update_nodes(force = false)
     sum = 0
     @nodes.each do |node|
-      sum += node.update_closest_centroid(@centroids)
+      sum += node.update_closest_centroid(@centroids, force)
     end
     sum
   end

lib/k_means/node.rb

@@ -17,10 +17,10 @@ def initialize(position, similarity_measure)
     @similarity_measure = similarity_measure
   end
 
-  def update_closest_centroid(centroids)
+  def update_closest_centroid(centroids, force = false)
     # If we haven't processed this node we need to give it an initial centroid
     # so that we have something to compare distances against
-    calculate_initial_centroid(centroids.first) unless @closest_centroid
+    calculate_initial_centroid(centroids.first) if (!@closest_centroid || force)  
 
     updated = false
     centroids.each do |centroid|
@@ -60,7 +60,7 @@ def calculate_distance(centroid)
     begin
       @position.send(@similarity_measure, centroid.position)
     rescue NoMethodError
-      raise "Hey, '#{@similarity_measure}' is not a measurement. Read the REAdME for available measurements"
+      raise "Hey, '#{@similarity_measure}' is not a measurement. Read the README for available measurements"
     end
   end
 

test/k_means/test_k_means.rb

@@ -1,6 +1,8 @@
 require 'helper'
 
+
 class TestKMeans < Test::Unit::TestCase
+
   context "A KMeans Instance" do
 
     setup do
@@ -20,31 +22,76 @@ class TestKMeans < Test::Unit::TestCase
       assert_equal 3, @kmeans.nodes.size
     end
 
+
   end
 
+  context "A Many tries" do
+    setup do
+      @kmeans2 = KMeans.new(data, :centroids => 2, :distance_measure => :euclidean_distance, :max_tries => 50)
+      @kmeans3 = KMeans.new(data, :centroids => 3, :distance_measure => :euclidean_distance, :max_tries => 50)
+    end
+
+    should "DBI of 3 clusters less than 2 clusters" do
+      assert @kmeans2.davies_bouldin_index > @kmeans3.davies_bouldin_index
+    end
+
+  end
+
+
   context "A KMeans Instance with specified initial centroids" do
     setup do
       @data = Array.new(3) {Array.new(2) {rand}}
-      class CustomCentroid
-        attr_accessor :position
-        def initialize(position); @position = position; end
-        def reposition(nodes, centroid_positions); end
-      end
 
-      @specified_centroids = @data[0..2].map { |d| CustomCentroid.new(d) }
-      @kmeans = KMeans.new(@data, :custom_centroids => @specified_centroids, :distance_measure => :cosine_similarity)
+      #@specified_centroids = @data[0..2].map { |d| CustomCentroid.new(d) }
+      @kmeans = KMeans.new(@data, :centroids => 3, :distance_measure => :cosine_similarity)
     end
 
     should "return an inspected array" do
-      assert_kind_of String, @kmeans.inspect
+      assert_kind_of(String, @kmeans.inspect)
     end
 
     should "have 3 centroids" do
-      assert_equal 3, @kmeans.centroids.size
-   end
+      assert_equal(3, @kmeans.centroids.size)
+    end
 
     should "have 3 nodes" do
-      assert_equal 3, @kmeans.nodes.size
+      assert_equal(3, @kmeans.nodes.size)
     end
   end
+
+  def data
+    [
+      [ 1.0, 1.0 ],
+      [ 2.0, 1.0 ],
+      [ 2.0, 1.0 ],
+      [ 1.5, 1.0 ],
+      [ 1.8, 1.5 ],
+      [ 1.2, 1.4 ],
+      [ 1.9, 2.0 ],
+      [ 1.9, 1.0 ],
+      [ 1.0, 1.9 ],
+      [ 2.0, 2.0 ],
+      [ 4.3, 1.1 ],
+      [ 4.8, 1.2 ],
+      [ 4.3, 1.5 ],
+      [ 4.4, 1.2 ],
+      [ 4.1, 1.9 ],
+      [ 4.8, 1.7 ],
+      [ 4.2, 1.1 ],
+      [ 4.4, 1.7 ],
+      [ 5.8, 5.2 ],
+      [ 5.2, 6.2 ],
+      [ 5.5, 6.2 ],
+      [ 5.1, 6.2 ],
+      [ 5.2, 6.3 ],
+      [ 5.8, 6.1 ],
+      [ 5.6, 6.3 ],
+      [ 5.3, 6.4 ],
+      [ 5.5, 6.1 ],
+      [ 5.1, 6.2 ],
+      [ 5.9, 6.2 ],
+      [ 6.2, 5.1 ]
+    ]
+  end
+
 end