c7bf0b7c1a86d5eb3be2c722cf2cf746-AuthorFeedback.pdf

Revealingthisconnection11 allows explicit reasoning about what were previously implicit trade-offs people made when choosing a particular12 method. Oursecond maincontribution15 is an efficient sampling-based estimation algorithm that specifically targets global feature importance. Since the first two are harder to evaluate for real datasets (see R2), we focused on unambiguous quantitative29 metrics. Using state-of-the-art models, particularly those with heavy37 feature engineering, is largely orthogonal to our purpose.