thresholding

• This technique is often used when the evaluation metric is F-score
  • You need to do thresholding to determine if an example is in the positive or negative class
• This is hard when:
  • 1. The train and test distributions are different
  • 2. You introduce new models into your blend
    • cause each model introduces new biases & variances
  • Solution: use percentile thresholding