validmind.RobustnessDiagnosis

RobustnessDiagnosis

@tags('sklearn', 'model_diagnosis', 'visualization')

@tasks('classification', 'regression')

defRobustnessDiagnosis(datasets:List[validmind.vm_models.VMDataset],model:validmind.vm_models.VMModel,metric:str=None,scaling_factor_std_dev_list:List[float]=DEFAULT_STD_DEV_LIST,performance_decay_threshold:float=DEFAULT_DECAY_THRESHOLD):

Assesses the robustness of a machine learning model by evaluating performance decay under noisy conditions.

Purpose

The Robustness Diagnosis test aims to evaluate the resilience of a machine learning model when subjected to perturbations or noise in its input data. This is essential for understanding the model's ability to handle real-world scenarios where data may be imperfect or corrupted.

Test Mechanism

This test introduces Gaussian noise to the numeric input features of the datasets at varying scales of standard deviation. The performance of the model is then measured using a specified metric. The process includes:

  • Adding Gaussian noise to numerical input features based on scaling factors.
  • Evaluating the model's performance on the perturbed data using metrics like AUC for classification tasks and MSE for regression tasks.
  • Aggregating and plotting the results to visualize performance decay relative to perturbation size.

Signs of High Risk

  • A significant drop in performance metrics with minimal noise.
  • Performance decay values exceeding the specified threshold.
  • Consistent failure to meet performance standards across multiple perturbation scales.

Strengths

  • Provides insights into the model's robustness against noisy or corrupted data.
  • Utilizes a variety of performance metrics suitable for both classification and regression tasks.
  • Visualization helps in understanding the extent of performance degradation.

Limitations

  • Gaussian noise might not adequately represent all types of real-world data perturbations.
  • Performance thresholds are somewhat arbitrary and might need tuning.
  • The test may not account for more complex or unstructured noise patterns that could affect model robustness.