Wikipedia Vandalism Detection

Vandalism detection on the Wikipedia history with JRuby v9.1.0.0+.

The Wikipedia Vandalism Detection Gem uses the Weka Machine-Learning Library via the weka gem.

What You can do with it

parsing Wikipedia history pages to get edits and revisions
creating training and test ARFF files from the WVC-PAN-10 and the WVC-PAN-11 (See also http://pan.webis.de under category Wikipedia Vandalism Detection: CLEF 2010 & CLEF 2011)
calculating vandalism features for a Wikipedia page (XML) from the history dump
creating and evaluating a classifier with the created training ARFF file
classifing new instances of Wikipedia edits as 'regular' or 'vandalism'

Installation

Add this line to your application's Gemfile:

gem 'wikipedia-vandalism_detection'

And then execute:

$ bundle

Or install it yourself as:

$ gem install wikipedia-vandalism_detection

Usage

require 'wikipedia/vandalism_detection'

Configuration

To configure the system put a wikipedia-vandalism-detection.yml file in the config/ or lib/config/ directory.

You can configure:

A) the training and test corpora directories and essential input and output files

corpora:
  base_directory: /home/user/corpora

  training:
    base_directory: training
    annotations_file: annotations.csv
    edits_file: edits.csv
    revisions_directory: revisions

  test:
    base_directory: test
    edits_file: edits.csv
    revisions_directory: revisons

output:
  base_directory: /home/user/output_path
  training:
    arff_file: training.arff
    index_file: training_index.yml
  test:
    arff_file: test.arff
    index_file: test_index.yml

Evaluation outputs are saved under the output base directory path.

B) the features used by the feature calculator

features:
  - anonymity
  - biased frequency
  - character sequence
  - ...

C) the classifier type and its options and the number of cross validation splits for the classifier evaluation

classifier:
  type: Trees::RandomForest         # Weka classifier class
  options: -I 10 -K 0.5             # same as for Weka, for further classifier options see Weka-dev documentation
  cross-validation-fold: 5          # default is 10
  training-data-options: balanced   # default is unbalanced

training-data-options is used to resample the training dataset:

unbalanced is the default value and uses the original dataset
balanced uses random undersampling of the majority class
oversampled uses SMOTE oversampling (with percentage -p) and random undersampling (with minority/majority class balance -u)

Examples:

# 200% SMOTE oversampling with 300% random undersampling
training-data-options: oversampled -p 200 -u true 300

# default 100% SMOTE oversampling with 300% random undersampling
training-data-options: oversampled -u true 300

# 200% SMOTE oversampling with default full (100% minority/majority class balance)
# random undersampling
training-data-options: oversampled -p 200

# default 100% SMOTE oversampling without undersampling
training-data-options: oversampled -u false

Instead of the true option you can also use t, y and yes as well as their upper case pendants.

Examples

Create training and test ARFF file from configured corpus:

training_dataset = Wikipedia::VandalismDetection::TrainingDataset.build
test_dataset = Wikipedia::VandalismDetection::TestDataset.build

While creating the training and test datasets, for each a corpus file index is created into the configured index_file directory. To run the corpus file index creation manually use:

Wikipedia::VandalismDetection::TrainingDataset.create_file_index!
Wikipedia::VandalismDetection::TestDataset.create_file_index!

Parse a Wikipedia page content:

At the moment no namespaces are supported while parsing a page. So, the <page>...</page> tags should not be included in a namespaced xml tag!

xml = File.read(wikipedia_page.xml)
parser = Wikipedia::VandalismDetection::PageParser.new
page = parser.parse(xml)

# Work with revisions and edits from the page
page.revisions.each do |revision|
  puts revison.id
  puts revison.parent_id
end

page.edits.each do |edit|
  puts edit.new_revision.id
  puts edit.old_revision.id
end

Use a classifier of configured type:

Create the classifier:

classifier = Wikipedia::VandalismDetection::Classifier.new

Evaluation of the classifier against the configured training corpus:

# classifier.classifier_instance returns the weka classifier instance
evaluation = classifier.classifier_instance.cross_validate(folds: 10)
puts evaluation.class_details

Classify a new edit:

# Classification of a Wikipedia Edit or a feature set
# 'edit' is a Wikipedia::VandalismDetection::Edit, this can be built manually or by
# parsing a Wikipedia page content and getting its edits
# The returned confidence is a value between 0.0 and 1.0 were 0.0 means 'regular' and 1.0 means 'vandalism'
confidence = classifier.classify(edit)

feature_calculator = Wikipedia::VandalismDetection::FeatureCalculator.new
features = feature_calculator.calculate_features_for(edit)
confidence = classifier.classify(features)

Evaluate test corpus classification:

evaluator = classifier.evaluator
# or create a new evaluator
evaluator = Wikipedia::VandalismDetection::Evaluator.new(classifier)

performance_data = evaluator.evaluate_testcorpus_classification #default sample_count = 100
performance_data = evaluator.evaluate_testcorpus_classification(sample_count: 200)

# following attributes can be used for further computations
recall_values = performance_data[:recalls]           # recall values for e.g. x-values of PRC or y-values of ROC
precision_values = performance_data[:precisions]     # precision values for e.g. y-values of PRC
fp_rate_values = performance_data[:fp_rates]         # false positive rate values for e.g. x-values of ROC
area_under_curve_pr = performance_data[:pr_auc]      # computed from the precision and recall values
area_under_curve_ro = performance_data[:roc_auc]     # computed from the recall and fp-rate values
total_recall = performance_data[:total_recall]       # precison and recall values with maximum area (rectangle area)
total_precision = performance_data[:total_precision]

Get each features predictive value for analysis:

evaluator = classifier.evaluator
# or create a new evaluator
evaluator = Wikipedia::VandalismDetection::Evaluator.new(classifier)

analysis_data = evaluator.feature_analysis #default sample_count = 100
analysis_data = evaluator.feature_analysis(sample_count: 1000)

This returns a hash comprising all feature names as configured as keys and the threshold hashes as values.

{
  feature_name_1:
    {
      0.0 => {fp:… , fn:… , tp:… , tn:… },
      …,
      1.0 => {fp:… , fn:… , tp:… , tn:… }
    },
  …,
  feature_name_n:
    {
      0.0 => {fp:… , fn:… , tp:… , tn:… },
      …,
      1.0 => {fp:… , fn:… , tp:… , tn:… }
    },
}

Creating new Features:

You can define your own new Feature classes and use them by configuration in the config.yml.

Make sure to define the Feature class inside of the Wikipedia::VandalismDetection::Features module and to implement the calculate method (also refer to the Wikipedia::VandalismDetection::Features::Base class definition).

module Wikipedia
  module VandalismDetection
    module Features
      class MyNewFeature < Base
        def calculate(edit)
          super # ensures raising an error if 'edit' is not an Edit.

          # ...your implementation
        end
      end
    end
  end
end

While creating new Feature classes you should be aware of the following naming convention: The feature's name in the config.yml is the downcased name with spaces or dashes of the feature class name

E.g.:

    features:
      - my new feature
      - my-new-feature

both search for a Feature class with the name MyNewFeature.

Contributing

Fork it ( http://github.com/paulgoetze/wikipedia-vandalism_detection/fork )
Create your feature branch (git checkout -b my-new-feature)
Commit your changes (git commit -am 'Add some feature')
Push to the branch (git push origin my-new-feature)
Create new Pull Request