What is point-in-time correctness in feature stores?

hard

Answer

$When constructing training data, for each event at time t, use only feature values known BEFORE t.$
$Prevents leakage from future data into features.$
$\mathrm{valid}_{\mathrm{from}}\; + \;\mathrm{valid}_{\mathrm{to}}$
$\operatorname{max}(\mathrm{feature}\mathrm{timestamp}\; \le \;\mathrm{event}\mathrm{timestamp})$
$(3) Feature store must maintain event log, not just current snapshot.$
$Without PIT, training features reflect future information that won't exist at serve time.$

Check yourself — multiple choice

Random
For each event at t, use only features known before t; requires $\mathrm{valid}_{\mathrm{from}} / \mathrm{valid}_{\mathrm{to}}$ timestamps + as-of joins + event log; prevents future leakage
Just current
Not real

PIT correctness: as-of joins + event log; prevents future leakage.

#features#data-quality

215 interview questions in this topic.