Growing Instability: Classifying Crisis Reports

This is a top 10 solution to the "Growing Instability: Classifying Crisis Reports" challenge from datasciencechallenge.org

The solution outline:

• Use only the data from 2005 onwards for training
• Discard the labels with less than 30 examples
• Vectorize the texts, apply TF-IDF weights
• For each label train an SVM
• Use out-of-fold predictions for train for selecting the best threshold
• For topics present in test and not present in train do a simple keyword matching