Bow_research

Bow tutorial

• Don't use RandomForest -> not good for sparse high dimensional data
• Use Tfidf instead of just sparse word2vec
• Use AUC instead of accuracy (increases the false positive rate by changing the threshold - plots true positives against that -> area under curve is indicator)
  • can deal with unbalanced data
  • handels all possible thresholds at once
  • doensn't care about the value range
  • metric in sklearn
• Use linear classifier (like LogisticRegression) - its faster and better and can handle d>n
• The more features you have the better -> consider n-grams