This repository contains the official implementation of the Neurocomputing article:
“Intrinsic Dimensionality as a Model-Free Measure of Class Imbalance”
Çağrı Eser, Zeynep Sonat Baltacı, Emre Akbaş, and Sinan Kalkan.
Neurocomputing 674 (2026) 132938
Paper: https://doi.org/10.1016/j.neucom.2026.132938
Preprint: https://arxiv.org/abs/2511.10475
We recommend using Python 3.8+ and a virtual environment (e.g. conda).
Dependencies (for ID estimation) to be installed (via pip, conda, etc.) include:
torch >= 1.12.0and any appropriate version oftorchvisionnumpyscikit-dimensiontqdm
For dependencies of individual integrations, please consult the relative README file of the respective method under methods/.
.
├── docs/
├── methods/
│ ├── Bag-of-Tricks/
│ ├── BCL/
│ ├── DRO-LT/
│ ├── GLMC/
│ ├── logit_adjustment/
│ └── SURE/
├── utils/
│ └── id-estimation/
│ ├── id_cifar.py
│ ├── id_imagenet.py
│ ├── id_places.py
│ └── README.md
├── LICENSE
└── README.md
We provide scripts to estimate ID on CIFAR-LT, PlacesLT and ImageNet-LT datasets in the utils/id-estimation directory.
This directory contains code for using our ID-based method with multiple integrations:
- Bag of Tricks (Zhang et al., 2021)
- Logit Adjustment (Menon et al., 2021)
- DRO-LT (Samuel and Chechik, 2021)
- BCL (Zhu et al., 2022)
- GLMC (Du et al., 2023)
- SURE (Li et al., 2024)
Each method has a dedicated directory under methods/ with its own instructions and an ID.md file describing how to plug in our ID-based measure.
We release a subset of the models used in the paper.
| Dataset | Method | Imbalance Ratio | Top-1 Accuracy | Checkpoints and Logs |
|---|---|---|---|---|
| CIFAR-10-LT | GLMC + ID | 100 | 87.9 | link |
| 50 | 90.5 | link | ||
| CIFAR-100-LT | GLMC + ID | 100 | 58.0 | link |
| 50 | 62.8 | link | ||
| CIFAR-10-LT | SURE + RW + ID | 100 | 87.0 | link |
| 50 | 90.4 | link | ||
| CIFAR-100-LT | SURE + RW + ID | 100 | 57.7 | link |
| 50 | 62.7 | link |
| Dataset | Method | Backbone | Top-1 Accuracy | Checkpoints and Logs |
|---|---|---|---|---|
| Places-LT | BoT + ID | ResNet-152 | 43.4 | link |
| ImageNet-LT | BoT + ID | ResNet-10 | 42.9 | link |
| ImageNet-LT | GLMC + ID | ResNeXt-50 | 56.3 | link |
| ImageNet-LT | BCL + ID | ResNet-50 (90EP) | 56.5 | link |
| ImageNet-LT | BCL + ID | ResNeXt-50 (90EP) | 57.9 | link |
| ImageNet-LT | BCL + ID | ResNeXt-50 (180EP) | 58.2 | link |
If you would like to cite this work, please use:
@article{eser2026intrinsic,
title = {Intrinsic dimensionality as a model-free measure of class imbalance},
journal = {Neurocomputing},
volume = {674},
pages = {132938},
year = {2026},
issn = {0925-2312},
doi = {https://doi.org/10.1016/j.neucom.2026.132938},
url = {https://www.sciencedirect.com/science/article/pii/S0925231226003358},
author = {Cagri Eser and Zeynep Sonat Baltaci and Emre Akbas and Sinan Kalkan},
keywords = {Intrinsic dimension, Long-tailed visual recognition, Class imbalance, Long-tailed learning},
abstract = {Imbalance in classification tasks is commonly quantified by the cardinalities of examples across classes. This, however, disregards the presence of redundant examples and inherent differences in the learning difficulties of classes. Alternatively, one can use complex measures such as training loss and uncertainty, which, however, depend on training a machine learning model. Our paper proposes using data Intrinsic Dimensionality (ID) as an easy-to-compute, model-free measure of imbalance that can be seamlessly incorporated into various imbalance mitigation methods. Our results across five different datasets with a diverse range of imbalance ratios show that ID consistently outperforms cardinality-based re-weighting and re-sampling techniques used in the literature. Moreover, we show that combining ID with cardinality can further improve performance. Our code and models are available at https://github.com/cagries/IDIM.}
}
For questions and suggestions, please contact:
- Cagri Eser - cagri.eser@metu.edu.tr