7.4 Pros and Cons

Pros

• Model agnostic approach
  
• Can potentially provide more accurate explanations than additive effects only
• Plots can be easy to interpret in many cases

Cons

• Calculation complexity - need to calculate contributions for p(p+1)/2 variables and pairwise interactions
• Credibility issues - for small datasets, net contributions are subject to larger randomness in ranking
  
• Procedure is not based on formal statistical test of significance and relies on heuristics.
  • false positives
    
  • false negatives - more likely with small samples
    
• With many variables and interactions, plots can be complex with small contributions to instance prediction