9.4 Pros and Cons

Pros

• Model agnostic
  
• Interpretable explanation
  
• Local fidelity
  
• Great for text and image analysis

Cons

• No consensus for best lower dimensional representations for continuous and category predictors
• Multiple implementations, so may get different results based on package used
  
• Glass box model fit to black-box predictions, not actual data, so quality of local fit may be misleading
  
• Finding an appropriate neighborhood difficult with high dimensional data