A New Way to Predict Probability Distributions

How confident can we be in a machine learning model’s prediction? This question has been a prominent area of research over the last decade, and it has major implications in high-stakes machine learning applications such as finance and healthcare. While many classification models, particularly <a href="https://towardsdatascience.com/why-calibrators-part-1-of-the-series-on-probability-calibration-9110831c6bde" rel="noopener" target="_blank">calibrated</a> models, come with uncertainty quantification by predicting a probability distribution over target classes, quantifying uncertainty in regression tasks is much more nuanced. Amongst many proposed methods, quantile regression is one of the most popular because no assumptions are made about the target distribution. Until recently, the main disadvantage of quantile regression was that one model had to be trained per predicted quantile. For instance, in order to predict the 10th, 50th, and 90th quantiles of a target distribution, three independent models would need to be trained. Catboost has since addressed this issue with the <a href="https://catboost.ai/en/docs/concepts/loss-functions-regression#MultiQuantile:~:text=MultiQuantile,-%5Cdisplaystyle%5Cfrac%7B%5Csum" rel="noopener ugc nofollow" target="_blank">multi-quantile loss function</a> — a loss function that enables a single model to predict an arbitrary number of quantiles. <a href="https://towardsdatascience.com/a-new-way-to-predict-probability-distributions-e7258349f464">Website</a>