wikipedia-vandalism_detection 0.0.1

data/lib/wikipedia/vandalism_detection.rb

@@ -0,0 +1,30 @@
+require 'wikipedia'
+require 'wikipedia/vandalism_detection/version'
+require 'wikipedia/vandalism_detection/configuration'
+require 'wikipedia/vandalism_detection/exceptions'
+
+require 'wikipedia/vandalism_detection/text'
+require 'wikipedia/vandalism_detection/revision'
+require 'wikipedia/vandalism_detection/edit'
+require 'wikipedia/vandalism_detection/page'
+require 'wikipedia/vandalism_detection/page_parser'
+require 'wikipedia/vandalism_detection/revision_parser'
+
+require 'wikipedia/vandalism_detection/word_lists'
+require 'wikipedia/vandalism_detection/diff'
+require 'wikipedia/vandalism_detection/wikitext_extractor'
+require 'wikipedia/vandalism_detection/features'
+require 'wikipedia/vandalism_detection/feature_calculator'
+
+require 'wikipedia/vandalism_detection/instances'
+require 'wikipedia/vandalism_detection/training_dataset'
+require 'wikipedia/vandalism_detection/test_dataset'
+require 'wikipedia/vandalism_detection/classifier'
+require 'wikipedia/vandalism_detection/evaluator'
+
+require 'weka/classifiers/meta/one_class_classifier'
+require 'weka/classifiers/meta/real_ada_boost'
+require 'weka/classifiers/functions/lib_svm'
+require 'weka/classifiers/trees/balanced_random_forest'
+
+require 'weka/filters/supervised/instance/smote'

data/lib/wikipedia/vandalism_detection/algorithms.rb

@@ -0,0 +1,18 @@
+require 'active_support/concern'
+require 'wikipedia/vandalism_detection/algorithms/kullback_leibler_divergence'
+
+module Wikipedia
+  module VandalismDetection
+    module Algorithms
+      extend ActiveSupport::Concern
+
+      included do
+        def kullback_leibler_divergence(text_a, text_b)
+          divergence = KullbackLeiblerDivergence.new
+          divergence.of text_a, text_b
+        end
+      end
+
+    end
+  end
+end

data/lib/wikipedia/vandalism_detection/algorithms/kullback_leibler_divergence.rb

@@ -0,0 +1,69 @@
+require 'wikipedia/vandalism_detection/features/base'
+
+module Wikipedia
+  module VandalismDetection
+    module Algorithms
+
+      class KullbackLeiblerDivergence
+
+        # Returns the Symmetric Kullback-Leibler divergence with simple back-off of the given text's character
+        # distribution. For implementation details, see: http://staff.science.uva.nl/~tsagias/?p=185
+        def of(text_a, text_b)
+          text_a = text_a.encode('UTF-8', 'binary', invalid: :replace, undef: :replace, replace: '')
+          text_b = text_b.encode('UTF-8', 'binary', invalid: :replace, undef: :replace, replace: '')
+
+          return Features::MISSING_VALUE unless !!text_a.match(/[[:alnum:]]/) && !!text_b.match(/[[:alnum:]]/)
+
+          distribution_a = character_distribution(text_a)
+          distribution_b = character_distribution(text_b)
+
+          sum_a = distribution_a.values.inject(:+)
+          sum_b = distribution_b.values.inject(:+)
+
+          character_diff = (distribution_b.keys - distribution_a.keys)
+
+          epsilon = [(distribution_a.values.min / sum_a), (distribution_b.values.min / sum_b)].min * 0.001
+          gamma = 1 - character_diff.size * epsilon
+
+          # check if values sum up to 1.0
+          sum_a_diff = 1.0 - distribution_a.values.inject(0) { |sum, value| sum += (value / sum_a) }.abs
+          sum_b_diff = 1.0 - distribution_b.values.inject(0) { |sum, value| sum += (value / sum_b) }.abs
+
+          raise(Exception, "Text a distr. does not sum up to 1.0") if sum_a_diff > 9e-6
+          raise(Exception, "Text b distr. does not sum up to 1.0") if sum_b_diff > 9e-6
+
+          divergence = 0.0
+
+          distribution_a.each do |character, distribution|
+            prob_a = distribution / sum_a
+            prob_b = (distribution_b.has_key?(character) ? (gamma * (distribution_b[character] / sum_b)) : epsilon)
+
+            divergence += ((prob_a - prob_b) * Math.log(prob_a / prob_b))
+          end
+
+          divergence
+        end
+
+        private
+
+        # Returns a hash representing each character's distribution
+        def character_distribution(text)
+          distribution = {}
+          return distribution if text.empty?
+
+          characters = text.downcase.scan(/[[:alnum:]]/)
+
+          characters.each do |character|
+            if distribution.has_key?(character.to_sym)
+              distribution[character.to_sym] += 1
+            else
+              distribution[character.to_sym] = 1
+            end
+          end
+
+          Hash[distribution.map { |key, value| [key, value / characters.count.to_f] }]
+        end
+      end
+    end
+  end
+end

data/lib/wikipedia/vandalism_detection/classifier.rb

@@ -0,0 +1,186 @@
+require 'ruby-band'
+require 'active_support/core_ext/string'
+require 'fileutils'
+
+require 'wikipedia/vandalism_detection/configuration'
+require 'wikipedia/vandalism_detection/edit'
+require 'wikipedia/vandalism_detection/feature_calculator'
+require 'wikipedia/vandalism_detection/instances'
+require 'wikipedia/vandalism_detection/evaluator'
+
+module Wikipedia
+  module VandalismDetection
+    class Classifier
+
+      attr_reader :evaluator, :dataset
+
+      # Loads the classifier instance configured in the config file.
+      def initialize(dataset = nil)
+        @config = Wikipedia::VandalismDetection.configuration
+        @feature_calculator = FeatureCalculator.new
+        @classifier = load_classifier(dataset)
+        @evaluator = Evaluator.new(self)
+      end
+
+      # Returns the concrete classifier instance configured in the config file
+      # When you configured a Trees::RandomForest classifier you will get a Weka::Classifiers::Trees::RandomForest
+      # instance.
+      # This instance can be used for native function callings of the classifier class.
+      def classifier_instance
+        @classifier
+      end
+
+      # Classifies an edit or a set of features and returns the vandalism confidence by default
+      # If option 'return_all_params: true' is set, it returns a Hash of form
+      # { confidence => ..., class_index => ...}
+      #
+      # @example
+      #   # suppose you have a dataset with 2 feature or 'edit' as an instance of Wikipedia::VandalismDetection::Edit
+      #   classifier = Wikipedia::VandalsimDetection::Classifier.new
+      #   features = [0.45, 0.67]
+      #
+      #   confidence = classifier.classify(features)
+      #   confidence = classifier.classify(edit)
+      def classify(edit_or_features, options = {})
+        features = @config.features
+        param_is_features = edit_or_features.is_a?(Array) && (edit_or_features.size == features.count)
+        param_is_edit = edit_or_features.is_a? Edit
+
+        unless param_is_edit || param_is_features
+          raise ArgumentError, "Input has to be an Edit or an Array of feature values."
+        end
+
+        feature_values = param_is_edit ? @feature_calculator.calculate_features_for(edit_or_features) : edit_or_features
+        return -1.0 if feature_values.empty?
+
+        feature_values = feature_values.map { |i| i == Features::MISSING_VALUE ? nil : i }
+
+        dataset = Instances.empty
+        dataset.set_class_index(feature_values.count)
+        dataset.add_instance([*feature_values, Instances::VANDALISM])
+
+        instance = dataset.instance(0)
+        instance.set_class_missing
+
+        if @config.use_occ?
+          if @config.classifier_options =~ /#{Instances::VANDALISM}/
+            index = Instances::VANDALISM_CLASS_INDEX
+          else
+            index = Instances::REGULAR_CLASS_INDEX
+          end
+        else
+          index = Instances::VANDALISM_CLASS_INDEX
+        end
+
+
+        confidence = (@classifier.distribution_for_instance(instance).to_a)[index]
+
+        if options[:return_all_params]
+          class_index = @classifier.classify_instance(instance)
+          class_index = class_index.nan? ? Instances::NOT_KNOWN_INDEX : class_index.to_i
+          results = { confidence: confidence, class_index: class_index }
+        else
+          results = confidence
+        end
+
+        results
+      end
+
+      # Cross validates the classifier.
+      # Fold is used as defined in configuration (default is 10).
+      #
+      # @example
+      #   classifier = Wikipedia::VandalismDetection::Classifier.new
+      #   evaluation = classifier.cross_validate
+      #   evaluation = classifier.cross_validate(equally_distributed: true)
+      #
+      def cross_validate(options = {})
+        @evaluator.cross_validate(options)
+      end
+
+      private
+
+      # Loads the (Weka-) Classifier set in the Configuration
+      def load_classifier(dataset)
+        classifier_name = @config.classifier_type
+
+        raise ClassifierNotConfiguredError, "You have to define a classifier type in config.yml" unless classifier_name
+        raise FeaturesNotConfiguredError, "You have to configure features in config.yml" if @config.features.blank?
+
+        begin
+          "Weka::Classifiers::#{classifier_name}::Base".constantize
+        rescue
+          raise ClassifierUnknownError, "The configured classifier type '#{classifier_name}' is unknown."
+        end
+
+        classifier_class = "Weka::Classifiers::#{classifier_name}::Base".constantize
+        options = @config.classifier_options
+
+        puts "Loading classifier #{classifier_name} with options '#{options}'..."
+
+        if dataset.nil?
+          if @config.balanced_training_data?
+            puts "using BALANCED training dataset"
+            dataset = TrainingDataset.balanced_instances
+          elsif @config.unbalanced_training_data?
+            puts "using FULL (unbalanced) training dataset"
+            dataset = TrainingDataset.instances
+          elsif @config.oversampled_training_data?
+            puts "using OVERSAMPLED training dataset"
+            dataset = TrainingDataset.oversampled_instances
+          end
+        end
+
+        if @config.use_occ?
+          dataset.rename_attribute_value(dataset.class_index, one_class_index, Instances::OUTLIER)
+        elsif @config.classifier_type.match('Functions::LibSVM')
+          dataset = remove_regular_instances(dataset)
+        end
+
+        @dataset = dataset
+
+        begin
+          classifier = classifier_class.new do
+            set_data dataset
+            set_class_index dataset.class_index
+            set_options options if options
+          end
+
+          classifier
+        rescue => e
+          raise "Error while loading classifier: #{e.class}: #{e.message}"
+        end
+      end
+
+      def one_class_index
+        if @config.classifier_options =~ /#{Instances::VANDALISM}/
+          Instances::REGULAR_CLASS_INDEX
+        else
+          Instances::VANDALISM_CLASS_INDEX
+        end
+      end
+
+      # Returns the given dataset cleaned up the regular instances
+      def remove_regular_instances(dataset)
+        features = @config.features
+
+        vandalism_dataset = Core::Type::Instances::Base.new do
+          features.each { |name| numeric :"#{name.gsub(' ', '_')}" }
+          nominal :class, [Instances::VANDALISM]
+        end
+
+        dataset.to_a2d.each_with_index do |attributes, index|
+          class_value = Instances::CLASSES[dataset.instance(index).value(dataset.class_index).to_i]
+          vandalism_dataset.add_instance([*attributes, class_value]) if  class_value == Instances::VANDALISM
+        end
+
+        filter = Weka::Filters::Unsupervised::Attribute::Normalize.new
+        filter.data(vandalism_dataset)
+        vandalism_dataset = filter.use
+        vandalism_dataset.class_index = features.count
+
+        vandalism_dataset
+      end
+    end
+  end
+end

data/lib/wikipedia/vandalism_detection/configuration.rb

@@ -0,0 +1,321 @@
+require 'weka/classifiers/meta/one_class_classifier'
+require 'singleton'
+
+module Wikipedia
+  module VandalismDetection
+
+    require 'yaml'
+
+    def self.configuration
+      Configuration.instance
+    end
+
+    class Configuration
+      include Singleton
+
+      TRAINING_DATA_BALANCED = 'balanced'
+      TRAINING_DATA_UNBALANCED = 'unbalanced'
+      TRAINING_DATA_OVERSAMPLED = 'oversampled'
+
+      attr_reader :data,
+                  :features,
+                  :classifier_options,
+                  :classifier_type,
+                  :cross_validation_fold,
+                  :output_base_directory,
+                  :training_data_options
+
+      def initialize
+        config = DefaultConfiguration[DefaultConfiguration::DEFAULTS]
+        @config_from_file ||= config.load_config_file(config.config_file)
+
+        @data ||= (@config_from_file ? config.deep_merge(@config_from_file) : config)
+
+        @classifier_type = @data['classifier']['type']
+        @classifier_options = @data['classifier']['options']
+        @cross_validation_fold = @data['classifier']['cross-validation-fold']
+        @training_data_options = @data['classifier']['training-data-options']
+        @replace_missing_values = @data['classifier']['replace-missing-values'].to_s
+
+        @features = @data['features']
+        @output_base_directory = File.expand_path(@data['output']['base_directory'], __FILE__)
+        @training_arff_file_name = @data['output']['training']['arff_file']
+        @test_arff_file_name = @data['output']['test']['arff_file']
+      end
+
+      # Returns whether the classifier uses one class classification
+      def use_occ?
+        @classifier_type == Weka::Classifiers::Meta::OneClassClassifier.type
+      end
+
+      def replace_training_data_missing_values?
+        !!(@replace_missing_values =~ /(true|t|yes|y)/i)
+      end
+
+      # Returns a boolean value whether a balanced data set is used for classifier training.
+      # (balanced means: same number of vandalism and regular samples)
+      def balanced_training_data?
+        @training_data_options == TRAINING_DATA_BALANCED
+      end
+
+      # Returns a boolean value whether an unbalanced data set is used for classifier training.
+      # (unbalanced means: vandalism and regular samples are used as given in arff file)
+      def unbalanced_training_data?
+        @training_data_options == TRAINING_DATA_UNBALANCED || @training_data_options.nil? ||
+            (!balanced_training_data? && !oversampled_training_data?)
+      end
+
+      # Returns a boolean value whether a oversampled data set is used for classifier training.
+      # (oversampled means: a balanced dataset is enriched through vandalism instances
+      # if vandalism number is less than regular number)
+      def oversampled_training_data?
+        !@training_data_options.nil? && @training_data_options.include?(TRAINING_DATA_OVERSAMPLED)
+      end
+
+      # Returns a hash of the oversampled training data options.
+      # Allowed options are -p (-percent) and -u (-undersampling)
+      def oversampling_options
+        if oversampled_training_data?
+          params = @training_data_options.gsub(TRAINING_DATA_OVERSAMPLED, '').split('-')
+
+          percent_default = 100.0
+          undersampling_default = 100.0
+
+          percent_option = params.select { |param| param.match /(p\s|percentage\s)\d+/i }[0]
+          undersampling_option = params.select { |param| param.match /(u\s|undersampling\s)/i }[0]
+
+          percent = percent_option.nil? ? percent_default : percent_option.split.last.to_f
+          undersampling = undersampling_default
+
+          if undersampling_option
+            if !undersampling_option.match(/(true|t|yes|y)/i).nil?
+              undersampling_percentage = undersampling_option.split.last
+              undersampling = undersampling_percentage.nil? ? undersampling_default : undersampling_percentage.to_f
+            else
+              undersampling = 0.0
+            end
+          end
+
+          { percentage: percent, undersampling: undersampling }
+        else
+          {}
+        end
+      end
+
+      # Returns the path to the classification file.
+      # Automatically sub directories for classifier and training data options are added.
+      # Thus it results in <output base dir>/<classifier name>/<training data options>/<file name>
+      def test_output_classification_file
+        classifiction_file_name = @data['output']['test']['classification_file']
+        classifier_name = @classifier_type.split('::').last.downcase
+
+        File.join(@output_base_directory, classifier_name,
+                  @training_data_options.gsub(/\s+/, '_'), classifiction_file_name)
+      end
+
+      # Returns the training arff file name.
+      # The path is expanded by used classifier & options and is in the same directory as the classification file.
+      def training_output_arff_file
+        File.join(File.dirname(test_output_classification_file), @training_arff_file_name)
+      end
+
+      # Returns the test arff file name.
+      # The path is expanded by used classifier & options and is in the same directory as the classification file.
+      def test_output_arff_file
+        File.join(File.dirname(test_output_classification_file), @test_arff_file_name)
+      end
+
+      # Returns file/path string for corpora files/directories and output files
+      # after following schema: <corpus type>_<progress stage>_<file name>.
+      #
+      # Instead of 'corpora' the word 'corpus' is used for grammatical reasons.
+      #
+      # example:
+      #   training_corpus_edits_file()
+      #   test_output_index_file()
+      #
+      def method_missing(method_name, *args)
+        return instance_variable_get("@#{method_name}") if instance_variable_defined?("@#{method_name}")
+
+        file_path_parts = method_name.to_s.split('_')
+
+        if file_path_parts.count >= 4
+          corpus_type = file_path_parts[0]
+          progress_stage = file_path_parts[1]
+          file_path = file_path_parts[2..-1].join('_')
+
+          if progress_stage == 'corpus'
+            progress_stage = 'corpora'
+            path = File.join(@data[progress_stage]['base_directory'], @data[progress_stage][corpus_type]['base_directory'])
+          elsif progress_stage == 'output'
+            path = @output_base_directory
+          else
+            return super
+          end
+
+          relative_path = File.join(path, @data[progress_stage][corpus_type][file_path])
+          absolute_path = File.expand_path(relative_path, __FILE__)
+          instance_variable_set "@#{method_name}", absolute_path
+        else
+          super
+        end
+      end
+    end
+
+    # This class represents the default config which is merged with the customized config from config YAML file.
+    class DefaultConfiguration < Hash
+      DEFAULTS = {
+          "source"    => Dir.pwd,
+          'features'  => [
+              "anonymity",
+              "anonymity previous",
+              "all wordlists frequency",
+              "all wordlists impact",
+              "article size",
+              "bad frequency",
+              "bad impact",
+              "biased frequency",
+              "biased impact",
+              "blanking",
+              "character sequence",
+              "character diversity",
+              "comment length",
+              "comment biased frequency",
+              "comment pronoun frequency",
+              "comment vulgarism frequency",
+              "compressibility",
+              "copyedit",
+              "digit ratio",
+              "edits per user",
+              "emoticons frequency",
+              "emoticons impact",
+              "inserted size",
+              "inserted words",
+              "inserted character distribution",
+              "inserted external links",
+              "inserted internal links",
+              "longest word",
+              "markup frequency",
+              "markup impact",
+              "non-alphanumeric ratio",
+              "personal life",
+              "pronoun frequency",
+              "pronoun impact",
+              "removed size",
+              "removed words",
+              "removed all wordlists frequency",
+              "removed bad frequency",
+              "removed biased frequency",
+              "removed character distribution",
+              "removed emoticons frequency",
+              "removed markup frequency",
+              "removed pronoun frequency",
+              "removed sex frequency",
+              "removed vulgarism frequency",
+              "replacement similarity",
+              "reverted",
+              "revisions character distribution",
+              "sex frequency",
+              "sex impact",
+              "same editor",
+              "size increment",
+              "size ratio",
+              "term frequency",
+              "time interval",
+              "time of day",
+              "upper case ratio",
+              "upper case words ratio",
+              "upper to lower case ratio",
+              "user reputation",
+              "vulgarism frequency",
+              "vulgarism impact",
+              "weekday",
+              "words increment"
+          ],
+          "corpora" => {
+            "base_directory" => nil,
+            "training" => {
+                "base_directory"      => nil,
+                "edits_file"          => nil,
+                "annotations_file"    => nil,
+                "revisions_directory" => nil
+            },
+            "test" => {
+                "base_directory"      => nil,
+                "edits_file"          => nil,
+                "revisions_directory" => nil,
+                "ground_truth_file"   => nil
+            }
+          },
+          "output" => {
+              "base_directory" => File.join(Dir.pwd, 'build'),
+              "training" => {
+                  "arff_file" => 'training.arff',
+                  "index_file" => 'training_index.yml',
+              },
+              "test" => {
+                  "arff_file" => 'test.arff',
+                  "index_file" => 'test_index.yml',
+                  "classification_file" => 'classification.txt'
+              }
+          },
+          "classifier" => {
+              "type"    => nil,
+              "options" => nil,
+              "cross-validation-fold" => 10,
+              "training-data-options" => 'unbalanced',
+              "replace-missing-values" => nil
+          }
+      }
+
+      def source
+        DEFAULTS['source']
+      end
+
+      # Looks in two places for a custom config file:
+      # in <app_root>/config/ and in <app_root>/lib/config
+      def config_file
+        config_file_path = "config/config.yml"
+        root_file = File.join(source, config_file_path)
+        lib_file = File.join(source, "lib/#{config_file_path}")
+        first_parent_file = find_first_parent_path_for(File.expand_path(File.dirname(__FILE__)), config_file_path)
+
+        File.exist?(root_file) ? root_file : (File.exist?(lib_file) ? lib_file : first_parent_file)
+      end
+
+      def load_config_file(file)
+        if File.exists?(file) && file =~ /config\.yml/
+          YAML.load_file(file)
+        else
+          warn %Q{
+
+            Configuration file not found in
+            #{source}/config,
+            #{source}/lib/config directory
+            or any other parent path.
+
+            To customize the system, create a config.yml file.
+
+          }
+        end
+      end
+
+      private
+
+      def find_first_parent_path_for(start_path, file)
+        file_path = File.join(start_path, file)
+
+        if File.exists?(file_path)
+          file_path
+        else
+          if File.dirname(start_path) != start_path
+            find_first_parent_path_for(File.dirname(start_path), file)
+          else
+            start_path
+          end
+        end
+      end
+
+    end
+  end
+end