Share the quantile function between regression and classification

[thibaultcordier]

Is your feature request related to a problem? Please describe.
The function that calculates the quantile for regression and classification is not the same.

-> mapie.utils.compute_quantiles
-> mapie.conformity_scores.interface.BaseConformityScore.get_quantile

Describe the solution you'd like
Reuse one of them for the other task, so you only have to manage one function.

[LEDazzio01]

Hi @Valentin-Laurent, I'd like to work on this.

After investigating the two functions, here's my understanding:

• _compute_quantiles (in mapie/utils.py) — used by classification conformity scores (LAC, TopK, APS, RAPS). Computes (n+1)(1-α)/n quantiles with method="higher". Handles 2D and 3D arrays (recursing on the last axis for 3D).

• BaseConformityScore.get_quantile (in mapie/conformity_scores/interface.py) — used by regression conformity scores. Supports reversed (upper/lower), unbounded (infinite sets), NaN-aware (nanquantile), and applies a ceiling correction: ceil(α*(n+1))/n.

Both apply a similar finite-sample correction ((n+1) factor), but the regression version has additional features (directionality, unbounded sets, NaN handling).

Proposed approach: Unify by making _compute_quantiles call get_quantile under the hood, or refactor both into a single function in utils.py that supports the full parameter set (reversed, unbounded, axis, 3D arrays). The classification callers would use the simpler defaults.

Happy to wait for your feedback on the direction before writing code.