Does the Explanation Drift Before the Guarantee? Label-Free Early Warning on Evolving Graphs
Deployed models can fail silently as the world they describe evolves. Attribution drift has served as an early warning in tabular data; we ask whether, in graph streams, a model's own explanations can forewarn the loss of nominal conformal coverage. Doing so requires repairing the naive drift statistic twice. A vanishi...
Tomasz Szandała, Damian Rączkowski
· 0 citations