Imbalanced data isn’t just a classification problem

What it does The repository holds the official code for the ICML 2021 long-talk paper that introduced Deep Imbalanced Regression. It learns from real-world data where continuous target values—such as age, depth, or health scores—are unevenly distributed, and some regions of the target space may have almost no examples. The codebase is split into per-dataset subfolders covering computer vision, NLP, and healthcare benchmarks.

The interesting bit Instead of forcing continuous problems into categorical buckets, the work treats the target space as a continuum. It introduces Label Distribution Smoothing (LDS) and Feature Distribution Smoothing (FDS) to re-weight and calibrate models across the entire range, including values that are rare or missing in training.

Key highlights

• Formalizes Deep Imbalanced Regression (DIR) with five curated benchmarks.
• Provides LDS for kernel-smoothed label re-weighting and FDS for feature-space calibration.
• Covers single-value prediction (age, similarity, health scores) and dense prediction (depth).
• Ships with a hands-on Colab tutorial and reproduced model checkpoints.
• Modular structure: separate code folders for each dataset domain.

Caveats

• Dataset preparation, training, and evaluation details are fragmented across per-dataset subfolders.
• Integrating LDS or FDS into your own model requires manual wiring into the training loop and network definition.

Verdict A solid stop if you are training regressors on skewed continuous data and need more than classification hacks. If your targets are balanced or purely categorical, this is not your toolbox.