See the original issue here stan-dev/loo#270. We can do something very similar to the approach outlined in the original issue.

This is yet another useful tool we could add to the ArviZ ecosystem to generate more accurate LOO-CV and ELPD estimates when we have many large Pareto-k values.