Statistics, Machine Learning & Citizen Science: Physical Mechanisms

A data workflow can include feature extraction, labeled training sets, validation sets, probabilistic classification, uncertainty estimates, anomaly scores, and human review. Citizen scientists can contribute pattern recognition and candidate vetting through structured interfaces.

Machine-learning models learn statistical relationships in training data and then predict labels or properties for new data. Their output can inherit selection biases, labeling errors, or distribution shifts, so performance must be evaluated on representative independent data.

Automated pipelines classify transients, galaxies, spectra, and exoplanet candidates at survey scale. Human-assisted projects have also identified unusual objects that automated filters initially missed.

Models learn correlations present in training data. Class imbalance, selection bias, label error, and distribution shift can therefore create confident but wrong classifications.