Skip to main content

Key terms

H2O Model Analyzer uses several key terms across its documentation, and each, in turn, is explained in the sections below.

Decision Support​

Validate current predictive behavior and explanation or simulate alternate scenarios for better decision-making.

Explanation​

Allows a user to explore local explanations for the selected observation.

Counterfactuals Explanations​

Allows a user to explore the closest alternate explanations for the selected observation. Users can use such recommendations to reason the minimal changes needed to affect the predictive behavior. Counterfactuals are searched using proximity metrics (L1 distance) within the restricted space of the supplied data.

Adversarial Explanations​

Allows a user to explore the closest alternate explanation for the selected row where the predictive decision is different. Such examples are explored by generating synthetic data using adversarial perturbation (the smallest change to the features of the selected observation).

Adversarial Robustness​

Adversarial robustness is related to the problem of finding vulnerabilities in a predicted model using adversarial examples, small indistinguishable changes to the input data resulting in incorrect predictive behavior. This form of robustness helps measure a model's resilience against such inputs.

Dimensionality Reduction​

Is the process that reduces the number of input features in a dataset while retaining as much variation in the original dataset as possible. Model Analyzer uses such projection techniques to visualize the data in the interactive view.

Uniform Manifold Approximation and Projection (UMAP)​

Uniform Manifold Approximation and Projection is a non-linear dimensionality reduction technique by McInnes et al. that helps in visualizing and understanding large, high dimensional datasets.

Principal Component Analysis (PCA)​

Principal component analysis (PCA) is a linear dimensionality reduction technique that helps in visualizing and understanding large, high dimensional datasets.

Feedback