influence functions estimate the effect of removing a training point on a
model without the need to retrain. They are based on a first-order Taylor
approximation that is guaranteed to be accurate for sufficiently small changes
to the model, and so are commonly used to study the effect