| Database | Estimation Anchor | Learner | Dataset | Feature Scaling |
|---|---|---|---|---|
| LCDB 1.1 | ⌈ 16 ⋅ 2n/8 ⌉ | 32 | 265 | none, min-max, standardization |
📦 Project Root
├── 📂 analysis — Analysis notebooks / Reproduce all figures from the paper
├── 📂 dataset — The LCDB 1.1 datasets
├── 📂 experiments — Workflow for organising LCDB 1.1 computation
├── 📂 lcdb_function — Functions for computing the learning curves
├── 📂 metadata — Meta data about LCDB 1.1
└── LCDB11_demo.ipynb - showing how to use our database
First download the precomputed learning curves from 4TU.ResearchData and extract them to the folder dataset. Or you can use the demonstration in LCDB11_demo.ipynb to download the data automatically. Useful metadata can be found in folder metadata.
Note, we recommend using learning curves from validation sets, since the sets differ in each inner and outer split, to ensure a no data-leakage version. The demonstration can be found in LCDB11_demo.ipynb.
We have stored all the probabilistic outputs and / or scores (when available for the learner). However, these files are too large for us to host for the broader public. If you are interested in a particular metric that we do not include, please contact us and we can compute it and host it.
This work is licensed under a CC BY 4.0 - - Creative Commons Attribution 4.0 International License. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/.
This work is primarily based on the OpenML dataset platform and the Scikit-learn library.
If you find our work useful, please cite our paper:
@inproceedings{yan2025lcdb,
title = {LCDB 1.1: A Database Illustrating Learning Curves Are More Ill-Behaved Than Previously Thought},
author = {Yan, Cheng and Mohr, Felix and Viering, Tom},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
volume = {38},
year = {2025}
}