Replies: 1 comment 2 replies
-
@LacombeLouis @gmartinonQM perhaps you can help? Great work on MAPIE by the way, awesome stuff! |
Beta Was this translation helpful? Give feedback.
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
This may be a really silly one ... 😅
In the literature the formula for the$(1-\alpha)$ quantile is typically explicitly stated with sample correction: the $(1-\alpha)$ quantile is defined as the $\frac{\lceil (1 - \alpha) (n+1)\rceil}{n}$ -th smallest value of the empirical distribution of non-conformity scores (see, for example, here). Barber et al. (2020) denote this $(1-\alpha)$ quantile as $\hat{q}_{n,\alpha}^+\{v\}$ .
In the package I therefore initially made this explicit (see here) like so,
where
v
corresponds to the vector of nonconformity scores.But I noted that when I implement the quantile in this way, it doesn't hold that$\hat{q}^{-}(v) =- \hat{q}^{+}(-v)$ where $\hat{q}^{-}(v)$ denotes the $\alpha$ quantile. I checked how this is implemented in MAPIE and it turns out they do not explicitly implement the correction (see here).
I assume the sample correction is just done under the hood by$(1-\alpha)$ .
Statistics.quantile
and that my initial approach was therefore wrong? Since #18 I'm following the same approach as in MAPIE, just applyingStatistics.quantile
directly tov
for a given coverage rateAny thoughts/comments would be much appreciated!
Beta Was this translation helpful? Give feedback.
All reactions