問題文
A team receives a new 900 million row event table with 40 columns and no documentation. They must produce a first analysis within a day. Which order of steps gives them the most reliable footing?
選択肢
- Profile each column's type, null rate and value range first, then look at single-variable distributions, then at relationships between the columns that the profile showed to be usable.
- Compute a correlation matrix over all 40 columns and start from the pairs with the largest absolute coefficients, since that immediately points at the structure worth investigating and ranks the columns by how much each one contributes to it.
- Sample 1 000 rows and inspect them by eye.
- Fit a gradient-boosted model on all 40 columns immediately and read the feature importances, because the importances rank the columns by usefulness and this saves the time that would otherwise be spent profiling columns that turn out to carry no signal at all.