imbalanced-regression

Why we included this project

Many real-world prediction tasks return continuous values that are unevenly distributed, such as age, health scores, or depth, where common ranges are well represented and rare ones barely covered. A standard regressor trained on such skewed data tends to collapse toward the densely sampled middle, which is exactly the problem this repository addresses. It distills an ICML 2021 paper into practical training techniques: benchmark datasets for age estimation, text similarity, health scoring, and depth prediction, plus label and feature distribution smoothing that rebalance loss and feature statistics during training so the model covers the full target range. A tutorial notebook and per-dataset code make the methods fairly easy to adapt to your own skewed data, even though the repo is organized around reproducing the paper's results. Teams building continuous predictions from long-tail data will find the core ideas directly portable.