Skip to main content
Version: Next

Metrics: Adversarial similarity

Overview​

H2O Model Validation offers an array of metrics to understand an adversarial similarity test. Below, each metric is described in turn.

Metrics​

Similar-to-secondary probabilities​

The similar-to-secondary probabilities histogram displays the row probabilities of the primary and secondary dataset rows.

  • X-axis: Probabilities of being similar to the secondary dataset
  • Y-axis: The relative frequency of the probability intervals between 0 and 1
    • The relative frequency (Y-axis) refers to the percent-wise frequencies or number of occurrences of values in the datasets. H2O Model Validation performs this relative or percent-wise calculation to compare numbers between the primary and secondary datasets. Since H2O Model Validation expects each dataset to have a different size, making a fair comparison on direct frequencies is complex. Therefore, H2O Model Validation divides the number of occurrences by the length of the dataset.
note
  • In the adversarial similarity test, predictions are generated to determine how much a dataset’s rows differ from those belonging to the training dataset.
  • Higher values to the histogram's right indicate a greater difference between the primary dataset and the secondary dataset.
  • A balance of values on both sides of the histogram indicates a 1.0 AUC score (a perfect separation between the two datasets).

probability of belonging to Secondary Dataset

Dissimilarity score​

The dissimilarity score card displays the general area under the curve (AUC) value, highlighting the dissimilarity between the primary and secondary datasets.

dissimilarity-score

Feature importance​

The feature importance chart displays the top features contributing to a higher area under the curve (AUC) value, highlighting the dissimilarity between the primary and secondary datasets. The chart displays the features from top to bottom.

  • X-axis: Gain value (the feature's importance to the test model)
  • Y-axis: Dataset variables (features)
note

Clicking on the bar of a feature triggers the display of the following plot: Feature particial dependency plot (PDP).

feature-importance

Feature particial dependency plot (PDP)​

The feature particial dependency (PDP) plot displays the impact the different values of the selected feature have on the predicted values. The selected feature refers to the feature clicked on the Feature importance chart. The features partial dependency (PDP) plot only displays one single PDP line instead of one for each dataset.

  • X-axis: A variable's values
  • Y-axis: Probabilities of being similar to the secondary dataset
  • Average probability: Average probability for the whole concatenated dataset

feature-pdp.png

note

The following histogram and graph are available for the observed feature in the feature partial dependence plot (PDP):

To access either the histogram or graph, consider the following instructions when viewing the Feature partial dependence plot (PDP):

  1. Click Kebab menu.


    kebab-menu

Feature histogram​

The feature importance histogram displays the selected feature on the Feature importance chart.

  • X-axis: A variable's values
  • Y-axis: The relative frequency of the variable's values

feature-histogram.png

Feature missing ratios​

The feature feature missing ratios graph displays any missing ratio values of the selected feature in the primary and secondary dataset. The selected feature refers to the feature selected on the Feature importance chart.

  • X-axis: A variable's values
  • Y-axis: Missing ratio values

feature-missing-ratios.png

Shapley tab​

Shapley table​

The Shapley table contains adversarial model results from a Shapley perspective. Each row in the table represents a row of the secondary dataset (being compared to the primary dataset). Rows are ordered from the highest to lowest predictive value. In other words, rows higher in the table represent the rows of the primary dataset most dissimilar to the secondary dataset.

shapley-table.png

note

  • H2O Model Validation in the Shapley table displays the top highest (dissimilar) or lowest (similar) results (prediction values).

  • The first two columns of the Shapley table refer to the row's ID, and the target column, while the rest of the columns refer to the train columns of the model.

    NameDescription
    Row-Nr.Row ID
    Adv-IsSecondaryTarget column

  • Clicking on one row ID displays a global or local Shapley chart that highlights the top most global and local features increasing the average model predictions that drive the most dissimilarity between the datasets. To learn more, see Global/local Shapley.

Show similar​

By default, rows in the Shapley table are ordered from highest to lowest predictive values. To switch the default order and view rows from lowest to highest predictive values, consider the following instructions:

  1. In the Shapley table, select Similar.

Show dissimilar​

To revert to the default order of the rows in the Shapley table, where test rows are ordered from highest to lowest predictive values, consider the following instructions:

  1. In the Shapley table, select Dissimilar.

Download Shapley values​

To download the Shapley values of the secondary dataset, consider the following instructions when viewing the Shapley table:

  1. Click Download Shapley values for the whole secondary dataset.
note

The CSV file contains adversarial model results from a Shapley perspective for the secondary dataset.

Global/local Shapley​

The global or local Shapley chart displays the top global or local features increasing the average model predictions the most while driving the most dissimilarity between the datasets.

  • X-axis: Global/local Shapley values
  • Y-axis: Dataset variables
global-shapleylocal-shapley
Feedback