The Task

Today's topics:

• External Data
• Features
• Gradient Boosting Machines
• LightGBM
• Stacking
• Style Breach Detection
• Results

Data

• PAN
  • StackExchange forum
  • 3k train, 1.5k validation, 1.5k test
  • 0 to 3 breaches
• External Movies
  • Movie reviews from Amazon
  • 270k
  • 0 or 1 breach
• External StackExchange
  • 35 StackExchange sites
  • 50k
  • 0 to 2 breaches

Text Segmentation

• Sliding Window
• 1/3 overlap
• number-of-words/3 segment size
• Max diff of feature vectors

Features (1)

• spaces
• digits
• commas
• (semi)colons
• apostrophes
• quotes
• parenthesis

• POS-tags
• short (< 4 chars)
• long (> 6 chars)
• average length
• all-caps
• capitalized

• question
• period
• exclamation
• short (<100chars)
• long (>200 chars)

• number of paragraphs

Characters:

Words:

Sentences:

Features (2)

• Stop words: you, the, is, of
• Function words: least, well, etc, whether
• Readability, e.g  Flesch reading ease:

• Vocabulary richness

• British/American English spelling

, f is a frequency function

Gradient Boosting Machines

LightGBM

• Leaf-wise (Best-first) Tree Growth
• Gradient-based One-Side Sampling
• Exclusive Feature Bundling
• Optimizations in Speed and Memory Usage

LightGBM + TF-IDF = 

• Character [2-6]-grams (up to 300k)
• Word [1-2]-grams (up to 300k)
• LogisticRegression for feature selection
• Parameter tuning
• Bagging

• Train TF-IDF on test set???

Top features

Stacking

Validation Results

Test Results

Style Breach Detection

• PAN 2017
• 134 training examples
• 0 to 8 breaches
• unbalanced classes (80:20)
• recursive search for breaches
• threshold > 50% ?

Results

Our CV on train results:

Last year PAN test results: