(ML Spec)

error analysis is the process of manually looking through a random sample of incorrectly predicted examples from cross-validation set and try to find out their commonalities (properties, themes, etc)

  • Some categories may overlap
  • Focus on category with most examples for improvements, for example
    • collect more data on this category
    • create more features related to this category

Caveats

  • error analysis is harder for tasks that humans are NOT good at
  • focus on the more promising things to try