Investigating Machine Learning Techniques to Improve Spec Tests III

Investigating Machine Learning Techniques to Improve Spec Tests III

Intro

This is a part of the series of blog posts related to Artificial Intelligence Implementation. If you are interested in the background of the story or how it goes:

This week we'll showcase training procedures for implementing Machine Learning models for general testing purposes. We will be using SerpApi's Google Organic Results Scraper API for the data collection. Also, you can check in the playground in more detailed view on the data we will use.

image


Breakdown of Custom Weighted-KNN in Ruby

Let's initialize a training class

class Train
  def initialize csv_path
    @@csv_path = csv_path
    @@vector_arr = []
    @@word_arr = []
    @@maximum_word_size = 100
    @@weights = Vector[]
    @@losses = []
  end
VariableExplanation
@@csv_pathThe path to key-specific CSV
@@vector_arrAn array of vectors representing tokenized words
@@word_arrAn array containing words with their corresponding types (0 or 1)
@@maximum_word_sizeMaximum allowed word size
@@weightsVector containing weights
@@lossesArray containing losses throughout training process

Let's read the key-specific CSV file.

def self.read
  @@word_arr = CSV.read(@@csv_path)
  @@word_arr
end

To give an example, if we type the following command:

csv_path = "organic_results/organic_results__snippet.csv"
Train.new csv_path
key_array = Train.read

The output would be:

[["0", "McDonald's: Burgers, Fries & More. Quality Ingredients."],
 ["0", "https://www.mcdonalds.com/us/en-us.html"],
 ["0", "https://www.mcdonalds.com › en-us"],
 ["1", "McDonalds.com is your hub for everything McDonald's. Find out more about our menu items and promotions today!"],
 ["0", "Breakfast"],
 ...

This array is a replicate of what is on the csv. It contains elements and their corresponding binary correspondence with the key in question.


Let's define the vectorized versions of these words within the Train class.

def self.define_training_set vectors
  @@vector_arr = vectors
end

To make such a vectors array, we'll use our vocabulary such as below:

{
  "<unk>": 0,
  "<pad>": 1,
  "1": 2,
  "mcdonald": 3,
  "'": 4,
  "s": 5,
  "burgers,": 6,
  "fries": 7,
  ...
}

Each of these token should be expressed in a vector with the following lines of command:

vector_array = key_array.map { |word| Database.word_to_tensor word[1] }
Train.define_training_set vector_array

Database.word_to_tensor is a function we defined in previous week's code. It tokenizes words or sentences in accordance with vocabulary. We actually added the following function within Database class to call for an already created vocabulary:

  def self.read_vocab vocab_path
    vocab = File.read vocab_path
    @@vocab = JSON.parse(vocab)
  end

The end result is an array of vectorizable arrays:

[[2],
 [3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
 [22, 23, 24, 10, 25],
 [22, 23, 24, 10, 30, 31, 32],
 [36, 10, 37, 38, 39, 40, 41, 42, 3, 4, 43, 44, 45, 9, 46, 47, 48, 49, 50, 51, 52, 53],
 [74],
 [22, 23, 24, 10, 75],
 [77, 78],
 [22, 23, 24, 10, 80],
 [82, 83],
 [22, 23, 24, 10, 85],
 ...
]

Let's define a function to automatically take the maximum word size to determine padding pattern. This is not recommended for the final model you want to train since the maximum in this set could be lesser than the actual example you'll feed.

def self.auto_define_maximum_size
  @@maximum_word_size = @@vector_arr.map {|el| el.size}.max
end

So to give an example again:

Train.auto_define_maximum_size

Output:

37

This is the word with maximum charachter count, so it'll be taken as the maximum word size. The rest of the items will be completed to 37 characters with <pad> character, tokenized as 1 in the vocabulary.


Let's define the function to extend a single vector to maximum word count.

def self.extend_vector vector
  vector_arr = vector.to_a
  (@@maximum_word_size - vector.size).times { vector_arr << 1 }
  Vector.[](*vector_arr)
end

It takes the vector, converts it to array, feed sufficient 1 into it, and then vectorizes it again. To go through the entire dataset we have:

def self.extend_vectors
  @@vector_arr.each_with_index do |vector, index|
    @@vector_arr[index] = extend_vector vector
  end
end

So if we apply the following command:

[Vector[2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
 Vector[3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
 Vector[22, 23, 24, 10, 25, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
 Vector[22, 23, 24, 10, 30, 31, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
 ...

We are greeted with multiple vectors each of which is sized 37. This is important in keeping the vectorial multiplication with weights possible in the following parts.


Now we need to initialize weights. They will help us tweak the vectors into the vicinity of their corresponding classification.

def self.initialize_weights
  weights = []
  @@maximum_word_size.times { weights << 1.0 }
  @@weights = Vector.[](*weights)
end

If we enter the following command, we can see that the output will be a vector of size 37, consisting of 1:

Train.initialize_weights

The output:

Vector[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]

Let's also add something to control changing parts:

def self.config k = 1, lr = 0.001
  [k, lr]
end
VariableExplanation
kThis is the number of neighbours we want to get the prediction of
lrThis is the number that defines the change in weights in each epoch

Let's define our function to tilt the vector in multiple dimensions to get the result we want.

def self.product vector
  @@weights.each_with_index do |weight, index|
    vector[index] = weight * vector[index]
  end

  vector
end

Each element within weight vector is multiplied by each element within the initial vector we use for training or prediction to get a weighted vector. Since our initial weight is consisting only of 1, This function will return the vector we feed into it in first iteration.

If our initial vector was consisting only of 2

vector = Vector[2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0]

Train.product vector

These commands would give us:

Vector[2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0]

Euclidean distance between two points in multidimensional space should not scare you, it's easy:

def self.euclidean_distance vector_1, vector_2
  subtractions = (vector_1 - vector_2).to_a
  subtractions.map! {|sub| sub = sub*sub }
  Math.sqrt(subtractions.sum)
end

You take the subtractions of each value within a vector, square them, sum all of them up and get the squareroot just like we mentioned in the previous blog post. So if we gave Vector[1,2], and Vector[2,3], it would calculate the distance between the points x=1, y=2 and x=2, y=3:

Train.euclidean_distance Vector[1,2], Vector[2,3]

Output:

1.4142135623730951

So to find the k nearest vectors to the given vector, we can plug in the distances:

def self.k_neighbors distances, k
  indexes = []
  (k).times do
    min = distances.index(distances.min)
    indexes << min
    distances[min] = distances.max + 1
  end

  indexes
end

There is another command responsible for collecting distances:

    distances = []
    @@vector_arr.each_with_index do |comparison_vector, vector_index|
      if vector_index == index
        distances << 100000000
      else
        distances << euclidean_distance(comparison_vector, vector)
      end
    end

    indexes = k_neighbors distances, k

It gives a big value to distance if the vector is itself to avoid confusion. If we plug the first vector, with k=1 for example, the output will be:

[22]

Which gives us the index of the closest vector to first vector.


Let's create a function for checking the classification of the indexes and giving us the mean of classifications.

def self.make_prediction indexes
  predictions = []
  indexes.each do |index|
    predictions << @@word_arr[index][0].to_i
  end

  predictions.sum/predictions.size
end

The outputs of such classifications could be 1 and 0 since we are checking if the element is snippet in this case. So their mean will be between 0 and 1, both ends included.

If we plug 22:

0

It shows us that the vector isn't snippet just like the vector at index 22.


Let's implement something to change the weights according to the validity of this prediction:

def self.update_weights result, indexes, vector, lr
  indexes.each do |index|
    subtractions = @@vector_arr[index] - vector
    subtractions.each_with_index do |sub, sub_index|
      if result == 0 && sub >= 0
        @@weights[sub_index] = @@weights[sub_index] + lr
      elsif result == 0 && sub < 0
        @@weights[sub_index] = @@weights[sub_index] - lr
      elsif result == 1 && sub >= 0
        @@weights[sub_index] = @@weights[sub_index] - lr
      elsif result == 1 && sub < 0
        @@weights[sub_index] = @@weights[sub_index] + lr
      end
    end
  end
end

result in here is 0 for false and 1 for true as the relationship between predicted classification, and real classification. We use lr to get the vectors closer, or further by adding or subtractiong lr from weights. Since the weighted vectors will be used for training, probabilistically, it won't do the same mistake with updated weights if the prediction is true, and will do a better job if the prediction is false. The result will be a change in weights:

Vector[0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999]

Notice that, not all of the weights might've changed in negative order to maximize, or minimize the distance between two vectors. It is a coincidence.


Mean Absolute Error is a pretty common technique for calculating loss:

def self.mean_absolute_error real, indexes
  errors = []
  indexes.each do |index|
    errors << (@@word_arr[index][0].to_i - real).abs
  end

  (errors.sum/errors.size).to_f
end

Calculating loss will tell us of the healthy training process. We might need to tweak the configuration to make it a better. Notice that it won't give indications as to if this is the process we must use, only the health of the process. You can find detailed information in previous blog posts.


Now let's cover the entire function for training one vector:

def self.train vector, index
    k, lr = config
    vector = extend_vector vector
    vector = product vector

    distances = []
    @@vector_arr.each_with_index do |comparison_vector, vector_index|
      if vector_index == index
        distances << 100000000
      else
        distances << euclidean_distance(comparison_vector, vector)
      end
    end

    indexes = k_neighbors distances, k
    real = @@word_arr[index][0].to_i
    prob_prediction = make_prediction indexes
    prediction = prob_prediction > 0.5 ? 1 : 0
    result = real == prediction ? 1 : 0

    update_weights result, indexes, vector, lr
    loss = mean_absolute_error real, indexes
    @@losses << loss

    puts "Result : #{real}, Prediction: #{prediction}"
    puts "Loss: #{loss}"

    prediction
end
  • We start by a vector, and its index.
  • We then call the k and lr values from config.
  • After that, we make the vector into a weighted vector.
  • Then we calculate the distance of the weighted vector with respect to every other vector in the training set.
  • Find the indexes of k closest neighbors.
  • Get the real classification of the weigthed vector.
  • Get the prediction of it using closest neighbors.
  • Compare them to create a result.
  • Update weights according to the results.
  • Calculate loss and append it to all losses. (For graphing it later)
  • Output the results and return the prediction.

If we do these steps iteratively, we will train the weights in a way that'll hold for the general case. Vector classified as 1 will get closer to vectors classified as 1, thus give us a good prediction.

Full Code


class Database
  def initialize json_data, vocab = { "<unk>" => 0, "<pad>" => 1 }
    super()
    @@pattern_data = []
    @@vocab = vocab
  end

  ## Related to creating main database
  def self.add_new_data_to_database json_data, csv_path = nil
    json_data.each do |result|
      recursive_hash_pattern result, ""
    end

    @@pattern_data = @@pattern_data.reject { |pattern| pattern.include? nil }.uniq.compact

    path = "#{csv_path}master_database.csv"
    File.write(path, @@pattern_data.map(&:to_csv).join)
  end

  def self.element_pattern result, pattern
    @@pattern_data.append([result, pattern].flatten)
  end

  def self.element_array_pattern result, pattern
    result.each do |element|
      element_pattern element, pattern
    end
  end

  def self.assign hash, key, pattern
    if hash[key].is_a?(Hash)
      if pattern.present?
        pattern = "#{pattern}__#{key}"
      else
        pattern = "#{key}"
      end

      recursive_hash_pattern hash[key], pattern
    elsif hash[key].present? && hash[key].is_a?(Array) && hash[key].first.is_a?(Hash)
      if pattern.present?
        pattern = "#{pattern}__#{key}__n"
      else
        pattern = "#{key}"
      end

      hash[key].each do |hash_inside_array|
        recursive_hash_pattern hash_inside_array, pattern
      end
    elsif hash[key].present? && hash[key].is_a?(Array)
      if pattern.present?
        pattern = "#{pattern}__n"
      else
        pattern = "#{key}"
      end

      element_array_pattern hash[key], pattern
    else
      if pattern.present?
        pattern = "#{pattern}__#{key}"
      else
        pattern = "#{key}"
      end

      element_pattern hash[key], pattern
    end
  end

  def self.recursive_hash_pattern hash, pattern
    hash.keys.each do |key|
      assign hash, key, pattern
    end
  end

  ## Related to tokenizing
  def self.default_dictionary_hash
    {
      /\"/ => "",
      /\'/ => " \'  ",
      /\./ => " . ",
      /,/ => ", ",
      /\!/ => " ! ",
      /\?/ => " ? ",
      /\;/ => " ",
      /\:/ => " ",
      /\(/ => " ( ",
      /\)/ => " ) ",
      /\// => " / ",
      /\s+/ => " ",
      /<br \/>/ => " , ",
      /http/ => "http",
      /https/ => " https ",
    }
  end

  def self.tokenizer word, dictionary_hash = default_dictionary_hash
    word = word.downcase

    dictionary_hash.keys.each do |key|
      word.sub!(key, dictionary_hash[key])
    end

    word.split
  end

  def self.iterate_ngrams token_list, ngrams = 1
    token_list.each do |token|
      1.upto(ngrams) do |n|
        permutations = (token_list.size - n + 1).times.map { |i| token_list[i...(i + n)] }

        permutations.each do |perm|
          key = perm.join(" ")

          unless @@vocab.keys.include? key
            @@vocab[key] = @@vocab.size
          end
        end
      end
    end
  end

  def self.word_to_tensor word
    token_list = tokenizer word
    token_list.map {|token| @@vocab[token]}
  end

  ## Related to creating key-specific databases 
  def self.create_key_specific_databases result_type = "organic_results", csv_path = nil, dictionary = nil, ngrams = nil, vocab_path = nil
    keys, examples = create_keys_and_examples

    keys.each do |key|
      specific_pattern_data = []
      @@pattern_data.each_with_index do |pattern, index|
        word = pattern.first.to_s

        next if word.blank?

        if dictionary.present?
          token_list = tokenizer word, dictionary
        else
          token_list = tokenizer word
        end

        if ngrams.present?
          iterate_ngrams token_list, ngrams
        else
          iterate_ngrams token_list
        end

        if key == pattern.second
          specific_pattern_data << [ 1, word ]
        elsif (examples[key].to_s.to_i == examples[key]) && word.to_i == word
          next
        elsif (examples[key].to_s.to_i == examples[key]) && word.numeric?
          specific_pattern_data << [ 0, word ]
        elsif examples[key].numeric? && word.numeric?
          next
        elsif key.split("__").last == pattern.second.to_s.split("__").last
          specific_pattern_data << [ 1, word ]
        else
          specific_pattern_data << [ 0, word ]
        end
      end

      path = "#{csv_path}#{result_type}__#{key}.csv"
      File.write(path, specific_pattern_data.map(&:to_csv).join)
    end

    if vocab_path.present?
      save_vocab vocab_path
    else
      save_vocab
    end
  end

  def self.create_keys_and_examples
    keys = @@pattern_data.map { |pattern| pattern.second }.uniq

    examples = {}
    keys.each do |key|
      examples[key] = @@pattern_data.find { |pattern| pattern.first.to_s if pattern.second == key }
    end

    [keys, examples]
  end

  def self.numeric?
    return true if self =~ /\A\d+\Z/
    true if Float(self) rescue false
  end

  def self.save_vocab vocab_path = ""
    path = "#{vocab_path}vocab.json"
    vocab = JSON.parse(@@vocab.to_json)
    File.write(path, JSON.pretty_generate(vocab))
  end

  def self.read_vocab vocab_path
    vocab = File.read vocab_path
    @@vocab = JSON.parse(vocab)
  end

  def self.return_vocab
    @@vocab
  end
end

class Train
  def initialize csv_path
    @@csv_path = csv_path
    @@vector_arr = []
    @@word_arr = []
    @@maximum_word_size = 100
    @@weights = Vector[]
    @@losses = []
  end

  def self.read
    @@word_arr = CSV.read(@@csv_path)
    @@word_arr
  end

  def self.define_training_set vectors
    @@vector_arr = vectors
  end

  def self.auto_define_maximum_size
    @@maximum_word_size = @@vector_arr.map {|el| el.size}.max
  end

  def self.extend_vector vector
    vector_arr = vector.to_a
    (@@maximum_word_size - vector.size).times { vector_arr << 1 }
    Vector.[](*vector_arr)
  end

  def self.extend_vectors
    @@vector_arr.each_with_index do |vector, index|
      @@vector_arr[index] = extend_vector vector
    end
  end

  def self.initialize_weights
    weights = []
    @@maximum_word_size.times { weights << 1.0 }
    @@weights = Vector.[](*weights)
  end

  def self.config k = 1, lr = 0.001
    [k, lr]
  end

  def self.product vector
    @@weights.each_with_index do |weight, index|
      vector[index] = weight * vector[index]
    end

    vector
  end

  def self.euclidean_distance vector_1, vector_2
    subtractions = (vector_1 - vector_2).to_a
    subtractions.map! {|sub| sub = sub*sub }
    Math.sqrt(subtractions.sum)
  end

  def self.k_neighbors distances, k
    indexes = []
    (k).times do
      min = distances.index(distances.min)
      indexes << min
      distances[min] = distances.max + 1
    end

    indexes
  end

  def self.make_prediction indexes
    predictions = []
    indexes.each do |index|
      predictions << @@word_arr[index][0].to_i
    end

    predictions.sum/predictions.size
  end

  def self.update_weights result, indexes, vector, lr
    indexes.each do |index|
      subtractions = @@vector_arr[index] - vector
      subtractions.each_with_index do |sub, sub_index|
        if result == 0 && sub >= 0
          @@weights[sub_index] = @@weights[sub_index] + lr
        elsif result == 0 && sub < 0
          @@weights[sub_index] = @@weights[sub_index] - lr
        elsif result == 1 && sub >= 0
          @@weights[sub_index] = @@weights[sub_index] - lr
        elsif result == 1 && sub < 0
          @@weights[sub_index] = @@weights[sub_index] + lr
        end
      end
    end
  end

  def self.mean_absolute_error real, indexes
    errors = []
    indexes.each do |index|
      errors << (@@word_arr[index][0].to_i - real).abs
    end

    (errors.sum/errors.size).to_f
  end

  def self.train vector, index
    k, lr = config
    vector = extend_vector vector
    vector = product vector

    distances = []
    @@vector_arr.each_with_index do |comparison_vector, vector_index|
      if vector_index == index
        distances << 100000000
      else
        distances << euclidean_distance(comparison_vector, vector)
      end
    end

    indexes = k_neighbors distances, k
    real = @@word_arr[index][0].to_i
    prob_prediction = make_prediction indexes
    prediction = prob_prediction > 0.5 ? 1 : 0
    result = real == prediction ? 1 : 0

    update_weights result, indexes, vector, lr
    loss = mean_absolute_error real, indexes
    @@losses << loss

    puts "Result : #{real}, Prediction: #{prediction}"
    puts "Loss: #{loss}"

    prediction
  end
end


json_path = "organic_results/example.json"
json_data = File.read(json_path)
json_data = JSON.parse(json_data)

Database.new json_data
## For training from scratch                     
#Database.add_new_data_to_database json_data, csv_path = "organic_results/"
#Database.create_key_specific_databases result_type = "organic_results", csv_path = "organic_results/", ngrams = 2
##
Database.read_vocab "vocab.json"

## We will use an iteration of csvs within a specific path in the end
csv_path = "organic_results/organic_results__snippet.csv"

Train.new csv_path
key_array = Train.read

vector_array = key_array.map { |word| Database.word_to_tensor word[1] }
Train.define_training_set vector_array
Train.auto_define_maximum_size
Train.extend_vectors
Train.initialize_weights
Train.config k = 3

vector_array.each_with_index do |vector, index|
  Train.train vector, index
end

Conclusion

From now on, this blog post series will be bi-weekly. Two weeks later, we will train the initial models, and showcase how to store them for implementation.

The end aim of this project is to create an open-source gem to be implemented by everyone using a JSON Data Structure in their code.

I'd like to thank the reader for their attention, and the brilliant people of SerpApi creating wonders even in times of hardship, and for all their support.

Join us on Dev.to | Hashnode | Twitter | YouTube