Principle:Scikit learn contrib Imbalanced learn Benchmark Dataset Loading
| Knowledge Sources | |
|---|---|
| Domains | Machine_Learning, Benchmarking, Imbalanced_Learning |
| Last Updated | 2026-02-09 03:00 GMT |
Overview
A standardized approach for loading curated benchmark datasets with known imbalance ratios to evaluate and compare resampling and classification methods.
Description
Benchmark dataset loading provides reproducible access to a curated collection of real-world imbalanced classification datasets. These datasets span various domains (biology, digit recognition, satellite imagery, etc.) with imbalance ratios ranging from 8.6:1 to over 40:1. Using standardized benchmarks ensures fair comparison across methods and reproducibility of experimental results.
The collection was compiled from UCI Machine Learning Repository datasets and is hosted on Zenodo for reliable access.
Usage
Use this principle when evaluating resampling techniques or ensemble classifiers on standardized benchmarks. Select datasets by name or ID to control which benchmarks are included in the evaluation.
Theoretical Basis
Benchmark evaluation follows the standard protocol:
- Load curated datasets with known imbalance ratios
- Split into train/test sets (stratified to preserve imbalance)
- Apply resampling or ensemble method
- Evaluate with imbalance-appropriate metrics (geometric mean, IBA, balanced accuracy)