Home Artificial Intelligence Probabilistic ML with Quantile Matching: an Example with Python Quantile regression Quantile matching

Probabilistic ML with Quantile Matching: an Example with Python Quantile regression Quantile matching

0
Probabilistic ML with Quantile Matching: an Example with Python
Quantile regression
Quantile matching

A little bit-known technique for turning quantile regression predictions right into a probability distribution.

Towards Data Science
“Quantile Matching”, by Giulia Roggia. Used with permission.

After we train regressive models, we obtain point predictions. Nonetheless, in practice we are sometimes considering estimating the uncertainty associated to every prediction. To attain that, we assume that the worth we try to predict is a random variable, and the goal is to estimate its distribution.

There are a lot of methods available to estimate uncertainty from predictions, corresponding to variance estimation, Bayesian methods, conformal predictions, etc. Quantile regression is one in every of these well-known methods.

Quantile regression consists in estimating one model for every quantile you might be considering. This could be achieved by means of an asymmetric loss function, often known as pinball loss. Quantile regression is easy, easy to grasp, and available in high performing libraries corresponding to LightGBM. Nonetheless, quantile regression presents some issues:

  • There isn’t a guarantee that the order of the quantiles can be correct. For instance, your prediction for the 50% quantile could possibly be greater than the one you get for the 60% quantile, which is absurd.
  • To acquire an estimate of your complete distribution, it’s good to train many models. As an illustration, should you need an estimate for every point percent quantile, you could have to coach 99 models.

Here’s how quantile matching might help.

The goal of quantile matching is to suit a distribution function given a sample of quantile estimates. We will frame this as a regression problem, so the curve doesn’t need to perfectly fit the quantiles. As an alternative, it must be “as close as possible”, while keeping the properties which make it a distribution function.

Specifically, we’re considering estimating the inverse cumulative distribution function: given a…

LEAVE A REPLY

Please enter your comment!
Please enter your name here