Political Party, Gender, and Age Classification Based on Political Blogs

1
Political Party, Gender, and Age Classification
Based on Political Blogs

• Michelle Hewlett and Elizabeth Lingg

2
Introduction

• Can individuals be classified by their writing
  style?
• Do people under 25 use different punctuation than
  those over 25?
• Do they use different words and phrases?
• Can you figure out someones political ideologies
  by analyzing their writing using probabilistic
  methods?

3
Classifier

• Hold Out Cross Validation
• 80 of Data in Training Set
• 20 of Data in Test Set
• Classify Bloggers using a Feature Vector
• Features generated from training data

4
Features

• Most frequent unigrams, bigrams, trigrams
• Bush, troops in Iraq, McCain
• Sentence length, Word length
• Punctuation
• Pronoun usage

5
Features

• Compute feature probabilities based on frequency
  in the training data
• If women use the word myself three times as
  often as men use the word myself,
  P(femalemyself) 75
• Pick features which are not 50/50 male/female or
  50/50 Republican/Democrat

6
Classification

• Using the feature vector to classify, bloggers
  with a low probability of being a Republican were
  classified as Democrat
• Writers with high Probability of being a
  Republican were classified as Republican
• Writers with moderate Probability were not
  classified or Unknown

7
Classifier Results
8
Classifier Results
9
Classifier Results
10
Clustering

• K-means clustering algorithm used with entire
  data set
• Used sum of absolute differences instead of
  Euclidean distance because our differences were
  so small
• Initialized centroids to a reasonable guess

11
Clustering Results
o Democrat Cluster 1 Democrat Cluster 2 o
Republican Cluster 1 Republican Cluster 2 o
Unknown Cluster 1 Unknown Cluster 2
12
Clustering Results
o Male Cluster 1 Male Cluster 2 o Female
Cluster 1 Female Cluster 2 o Unknown Cluster
1 Unknown Cluster 2
13
Conclusion

• It is possible to identify the characteristics of
  a writer based on writing style, words and
  phrases!
• Political Party gave the best results, followed
  by Gender, then Age

14
Future Work

• Generalize results with a larger data set and
  greater number of features
• Generalize results in a different domain
• Possibly implement linear regressions, logistic
  regressions, SVM