class LrLinearRegression

RubyLinearRegression

Attributes

lambda[R]

mu[R]

normalize[R]

sigma[R]

theta[R]

x[R]

y[R]

Public Class Methods

new() click to toggle source

# File lib/lr_linear_regression.rb, line 8
def initialize
  @mu = 0
  @sigma = 1
end

Public Instance Methods

compute_cost(test_x = nil, test_y = nil) click to toggle source

Compute the mean squared cost / error function

# File lib/lr_linear_regression.rb, line 35
def compute_cost test_x = nil, test_y = nil

  if not test_x.nil?
    test_x.each_index do |row|
      test_x[row].each_index do |i|
        test_x[row][i] = (test_x[row][i] - @mu[i]) / @sigma[i].to_f
      end
    end if @normalize
    test_x = test_x.map { |r| [1].concat(r) }
  end

  # per default use training data to compute cost if no data is given
  cost_x = test_x.nil? ? @x : Matrix.rows( test_x )
  cost_y = test_y.nil? ? @y : Matrix.rows( test_y.collect { |e| [e] } )

  # First use matrix multiplication and vector subtracton to find errors
  errors = (cost_x * @theta) - cost_y

  # Then square all errors
  errors = errors.map { |e| (e.to_f**2)  }

  # Find the mean of the square errors
  mean_square_error = 0.5 * (errors.inject{ |sum, e| sum + e }.to_f / errors.row_size)

  return mean_square_error
end

load_training_data(x_data, y_data, normalize = true) click to toggle source

Loads and normalizes the training data, must be called prior to training. Arguments:

x_data: (Two dimensiolnal array with the independent variables of your training data)
y_data: (Array with the dependent variables of your training data)

# File lib/lr_linear_regression.rb, line 17
def load_training_data x_data, y_data, normalize = true

      @normalize = normalize

      # normalize the x_data
      x_data = normalize_data( x_data ) if @normalize

      # add 1 column to our data
      x_data = x_data.map { |r| [1].concat(r) }

      # build our x Matrix & y Vector
      @x = Matrix.rows( x_data )
      @y = Matrix.rows( y_data.collect { |e| [e] } )

      @theta = Matrix.zero(@x.column_size, 1)
end

predict(data) click to toggle source

Makes a prediction based on your trained model. train_normal_equation must be called prior to making a prediction.

Arguments:

data: (Array of independent variables to base your prediction on)

# File lib/lr_linear_regression.rb, line 102
def predict data

  # normalize
  data.each_index do |i|
    data[i] = (data[i] - @mu[i]) / @sigma[i].to_f
  end if @normalize

  # add 1 column to prediction data
  data = [1].concat( data )

  # perform prediction
  prediction = (Matrix[data] * @theta)[0,0].to_f

  return prediction

end

train_gradient_descent( alpha = 0.01, iterations = 500, verbose = false ) click to toggle source

Calculate optimal theta using gradient descent Arguments:

alpha: Learning rate
iterations: Number of iterations to run gradient descent
verbose: If true will output cost after each iteration, can be used to find optimal learning rate (alpha) and iteration

# File lib/lr_linear_regression.rb, line 82
def train_gradient_descent( alpha = 0.01, iterations = 500, verbose = false )

  0.upto( iterations ) do |i|
    @temp_theta = Array.new(@theta.row_size)
    0.upto(@theta.row_size-1) do |row|
      @temp_theta[row] = @theta[row,0] - alpha * compute_gradient(row)
    end

    @theta = Matrix.columns([@temp_theta])

    puts "Cost after #{i} iterations = #{compute_cost}" if verbose
  end

end

train_normal_equation(l = 0) click to toggle source

Calculate the optimal theta using the normal equation

# File lib/lr_linear_regression.rb, line 63
def train_normal_equation l = 0

  @lambda = l
  lambda_matrix = Matrix.build(@theta.row_size,@theta.row_size) do |c,r|
      (( c == 0 && r == 0) || c != r) ? 0 : 1;
    end

  # Calculate the optimal theta using the normal equation
  # theta = ( X' * X )^1 * X' * y
  @theta = (@x.transpose * @x + @lambda * lambda_matrix ).inverse * @x.transpose * @y

  return @theta
end

Private Instance Methods

compute_gradient( parameter ) click to toggle source

Compute the mean squared cost / error function

# File lib/lr_linear_regression.rb, line 148
def compute_gradient( parameter )

  # First use matrix multiplication and vector subtracton to find errors
  gradients = ((@x * @theta) - @y).transpose * @x.column(parameter)

  # Mean the grandient
  mean = gradients.inject{ |sum, e| sum + e } / gradients.size

  return mean
end

normalize_data(x_data, mu = nil, sigma = nil) click to toggle source

# File lib/lr_linear_regression.rb, line 120
def normalize_data(x_data, mu = nil, sigma = nil)

  row_size = x_data.size
  column_count = x_data[0].is_a?( Array) ? x_data[0].size : 1

  x_norm = Array.new(row_size)
  @mu = Array.new(column_count)
  @sigma = Array.new(column_count)

  0.upto(column_count - 1) do |column|
    column_data = x_data.map{ |e| e[column] }
    @mu[column] = column_data.inject{ |sum, e| sum + e } / row_size
    @sigma[column] = (column_data.max - column_data.min)
  end

  0.upto(row_size-1) do |row|
    row_data = x_data[row]
    x_norm[row] = Array.new(column_count)
    row_data.each_index do |i|
      x_norm[row][i] = (row_data[i] - @mu[i]) / @sigma[i].to_f
    end
  end

  return x_norm

end