問題文
A feature must express how much each transaction deviates from the customer's own typical amount. Which construction avoids using information the model would not have at prediction time?
選択肢
- Compute the customer's statistic over the transactions strictly before the one being scored, so each row is compared against its own past only.
- Rank the amounts within the whole table.
- Compute the customer's statistic over all their transactions in the table and subtract it from each amount, because the statistic describes the customer rather than any individual transaction and is therefore a property of the customer that is legitimately available at scoring time.
- Subtract the global average amount across all customers, so every transaction is measured against one figure that the model can be given at prediction time.