scikit-learn-contrib / MAPIE

A scikit-learn-compatible module to estimate prediction intervals and control risks based on conformal predictions.
https://mapie.readthedocs.io/en/latest/
BSD 3-Clause "New" or "Revised" License
1.3k stars 111 forks source link

Adaptative / Conditional Conformal Prediction #300

Open thibaultcordier opened 1 year ago

thibaultcordier commented 1 year ago

Is your feature request related to a problem? Please describe.

Despite attractive theoretical guarantees and practical successes, Predictive Interval (PI) given by Conformal Prediction (CP) may not reflect the uncertainty of a given model. This limitation arises from CP methods using a constant correction for all test points, disregarding their individual uncertainties, to ensure coverage properties. (From Salim I. Amoukou, N. Brunel (2023)).

Different methods are proposed to deal with this problem, which we have grouped under the name "adaptive conformal prediction" (see additional context).

Describe the solution you'd like

The objective of this enhancement is to propose conditional / adaptative conformal prediction for MapieRegressor and MapieClassifier (in two different issues/PRs ?).

Describe alternatives you've considered

One way to achieve this is to adapt the fit and predict methods to the different classes of ConformityScore.

Additional context

In "adaptive conformal prediction", we distinguish different approaches (non exhaustive list):

jesuinovieira commented 1 year ago

It seems that this issue was closed together with #305 🤔

thibaultcordier commented 1 year ago

Good point, thank you for this comment :)

joshdunnlime commented 1 year ago

What about Self-supervised Conformal Prediction (SSCP)? Is that likely to get added? +1 from me :)