Noise & Robustness Playground
Inject label noise, feature noise, missingness or outliers and measure how classification performance degrades.
What to observe while you experiment
Robustness measures how performance changes when the data are corrupted or shifted. Label noise, feature noise, missingness and outliers attack different parts of the learning problem, so degradation patterns reveal model and preprocessing sensitivity.
Experiment deliberately
Choose one model and sweep a single corruption type from zero upward. Predict whether the metric should degrade smoothly or fail abruptly and explain why.
Robustness is conditional. Corrupt training data to study learning under imperfect evidence, or corrupt evaluation data to study deployment-time degradation.
Ready.