Let a multi-armed bandit pick your active learning strategy

What it does

libact is a Python toolkit for pool-based active learning. It collects roughly a dozen query strategies—uncertainty sampling, core-set diversity, BALD, density-weighted methods, and others—behind a single interface so you can swap algorithms without rebuilding your pipeline. It also ships with lightweight built-in models and adapters that let you wrap scikit-learn estimators for use inside its active learning loop.

The interesting bit

The project’s standout feature is ActiveLearningByLearning, a meta-algorithm that treats each query strategy as an arm in a multi-armed bandit and switches among them dynamically. It is a rare built-in admission that the “best” strategy depends on the dataset and the moment, and that letting an algorithm shop around often beats committing upfront.

Key highlights

• Bundles 12+ query strategies spanning uncertainty, diversity, density, and disagreement-based selection.
• Includes a multi-armed bandit meta-strategy that auto-selects among query algorithms during a run.
• Offers SklearnAdapter and SklearnProbaAdapter to drop existing scikit-learn classifiers into the loop.
• Two strategies (VarianceReduction and HintSVM) require compiled C extensions; the rest are pure Python.
• Supports Python 3.9 through 3.12 and tracks recent numpy/scikit-learn releases.

Caveats

• Native Windows support is explicitly discouraged; the README recommends WSL instead.
• Installing from source requires BLAS/LAPACKE and a meson/ninja/cython toolchain.
• The C extensions for VarianceReduction and HintSVM add system-level build friction.

Verdict

Worth a look if you are benchmarking active learning methods or need a plug-and-play loop with strategy swapping. Skip it if you just need a single uncertainty sampler and already own the surrounding data pipeline.