OpenMOA

What is OpenMOA?

Real-world data streams rarely stay the same. Sensors fail, features appear and disappear, and the world keeps changing — yet most online learning libraries assume a fixed feature space. OpenMOA is built for the messy reality.

OpenMOA is a Python library for Utilitarian Online Learning (UOL): online learning under dynamic, evolving feature spaces. It provides:

• 10 state-of-the-art UOL algorithms — from sparse linear models to deep graph neural networks
• 5 stream wrappers — simulate every major type of feature-space evolution
• Full integration with CapyMOA — access 30+ classic stream learners, 12 drift detectors, and 40+ datasets out of the box
• Clean, consistent API — train, predict, and evaluate with the same interface across all algorithms

⚠️ Early Development OpenMOA is actively developed and the API may change before v1.0.0. If you run into issues, please open a GitHub Issue or reach out on Discord.

Installation

OpenMOA requires Java (for MOA backend) and PyTorch (for deep learning algorithms).

# 1. Check Java is installed
java -version

# 2. Install PyTorch (CPU version)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

# 3. Install OpenMOA
pip install openmoa

# 4. Verify
python -c "import openmoa; print(openmoa.__version__)"

Having trouble? See the full installation guide for GPU support and platform-specific instructions.

Quick Start

from openmoa.datasets import Electricity
from openmoa.classifier import AdaptiveRandomForestClassifier
from openmoa.evaluation import prequential_evaluation

stream = Electricity()
learner = AdaptiveRandomForestClassifier(schema=stream.get_schema())

results = prequential_evaluation(stream, learner, window_size=1000)

print(f"Accuracy:  {results.cumulative.accuracy():.2f}%")
print(f"Kappa:     {results.cumulative.kappa():.4f}")
print(f"Wall time: {results.wallclock():.2f}s")

Running a UOL algorithm on an evolving feature stream is just as straightforward:

from openmoa.stream import OpenFeatureStream
from openmoa.datasets import Electricity
from openmoa.classifier import OASFClassifier
from openmoa.evaluation import prequential_evaluation

# Simulate a pyramid feature-space evolution
stream = OpenFeatureStream(
    base_stream=Electricity(),
    evolution_pattern="pyramid",
    d_min=2,
    d_max=8,
    total_instances=45312,
)

learner = OASFClassifier(schema=stream.get_schema())
results = prequential_evaluation(stream, learner)
print(f"Accuracy: {results.cumulative.accuracy():.2f}%")

UOL Algorithms

OpenMOA introduces 10 original algorithms designed specifically for dynamic feature spaces. Each handles the challenge of features appearing, disappearing, or changing over time.

from openmoa.classifier import (
    FESLClassifier, OASFClassifier, RSOLClassifier,
    FOBOSClassifier, FTRLClassifier, OVFMClassifier,
    OSLMFClassifier, ORF3VClassifier, OLD3SClassifier, OWSSClassifier,
)

Stream Wrappers

Not sure which feature-evolution pattern fits your application? OpenMOA's stream wrappers let you simulate five distinct paradigms on any existing dataset:

from openmoa.stream import (
    OpenFeatureStream, TrapezoidalStream,
    CapriciousStream, EvolvableStream, ShuffledStream,
)

Benchmark

Our benchmark evaluates ten representative UOL and OL algorithms across 12 binary and 9 multi-class datasets under three dynamic feature-space paradigms. All experiments use OpenMOA's unified API with standardized prequential evaluation, ensuring fair and reproducible comparison.

The following performance results are from the latest code review and optimization pass (Windows 11, Intel CPU, Python 3.13, NumPy 2.x):

End-to-end throughput (400 training steps, d = 8 features):

Benchmark code: demo/demo_fesl_benchmark_binary.py

What Else is Inside?

OpenMOA is built on top of CapyMOA and inherits its full ecosystem:

• 30+ classic stream classifiers — Hoeffding Tree, ARF, EFDT, Naive Bayes, kNN, and more
• 12 concept drift detectors — ADWIN, DDM, HDDM, Page-Hinkley, ABCD (multivariate), and more
• Regression support — FIMTDD, ARFFIMTDD, SOKNL, ORTO, FESLRegressor, OASFRegressor
• Semi-supervised evaluation — prequential_ssl_evaluation with configurable label probability
• Online Continual Learning — ocl_train_eval_loop with forward/backward transfer metrics
• 40+ benchmark datasets — Electricity, Covtype, RCV1, Fried, Bike, and more

Cite Us

If you use OpenMOA in your research, please cite:

@misc{ZhiliWang2025OpenMOAAPythonLibraryforUtilitarianOnlineLearning,
    title={{OpenMOA}: A Python Library for Utilitarian Online Learning},
    author={Zhili Wang, Heitor M.Gomes and Yi He},
    year={2025},
    eprint={},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/},
}