Reproduce “regAL: Python Package for Active Learning of Regression Problems” on OpenResearch

About this paper

Elizaveta Surzhikova, Jonny Proppe

The `regAL` Python package facilitates active learning for regression problems, enabling efficient machine learning model development with minimal data labeling through various sample selection methods and robust benchmarking. It demonstrated that Covariance Sampling and Query by Committee achieve the lowest error progression on a QM9 dataset, outperforming random sampling.

Key insights

• Covariance Sampling and Query by Committee generally achieve superior performance in reducing model error for regression tasks compared to simpler or passive strategies.
• The detailed run history logging is crucial for understanding the active learning trajectory, analyzing how different strategies influence model development, and ensuring experiment reproducibility.
• The computational overhead of `regAL`'s active learning methods is generally negligible when compared to the typical time and resource expenditures associated with acquiring new labels, which is often the primary bottleneck.

Summary by alphaXiv.