WikiAutoEncoder

An auto-encoder to help classify article by category on Wikipedia.

The Task:

1. Take N classes of wikipedia articles (laziest way possible)
   • for each class - 1000 articles
2. Create an autoencoder to compress the articles.
3. Perform classification with typical classifier.
4. Discussion:
   • Compare to a classification on plain text.
   • Compare to PCA

What have I used for this task ?

• Crawling Wikipedia:
  • mediawiki API.
  • requests python package.
• Word / Article Representation:
  • stopwords removal from nltk.
  • co-occurance probability with a default context size of 6 for each token.
  • article representation through simple sum of co-occurance probability for each token.
• Compression:
  • AutoEncoder:
    • a simple 4 layers NN (implemented with tensorflow).
  • PCA:
    • sklearn implementation of PCA.
• Classification:
  • LogisticRegression with sklearn pipeline.
  • RandomForest with sklearn pipeline.
• Visualization:
  • Visualization using matplotlib.