Continuous Testing Feedback and Observability
Once a model is in production, Robust Intelligence can provide detailed information on the model’s performance to enable you to identify and correct issues.
Model monitors
A model under Continuous Testing displays summary information about model health on the Overview page of the project that contains the model. Robust Intelligence monitors machine learning model performance across the following three risk categories:
Operational tests a model’s overall performance and data stability of over time.
Security tests a model’s resilience against compromise from external attacks.
Fairness tests a model’s outcome for fair treatment among subsets in the data.
Viewing the CT risk pages
Sign in to a RI Platform instance.
The Workspaces page appears.
Click a workspace.
The Workspaces summary page appears.
Select a project.
You can filter or sort the list of projects in a workspace with the Sort and Filter controls in the upper right. Click the glyph to the right of the Filter control to switch between list and card display for projects. Type a string in Search Projects… to display only projects that match the string.
Select the Continuous Testing tab in the left panel.
The CT Overview page appears
Click on one of the risk category tabs at the top (Operational, Security, or Fairness).
The corresponding CT risk page appears.
Monitors and risk categories
Monitors of particular interest can be pinned to the top of the list by clicking the pushpin icon at the top left corner of the monitor. Pinned Operational risk monitors will also show up in the Continuous Testing Overview page. Use the Date Range controls to change your view of the monitor chart. Some monitors have feature or subset dropdown menus which allow you to view the monitored metric for a specific feature or subset(s) of a feature in your dataset.
Enabling or disabling notifications for a monitor
You can enable or suppress notifications from a specified monitor.
Sign in to a RI Platform instance.
The Workspaces page appears.
Click a workspace.
The Workspaces summary page appears.
Select a project.
You can filter or sort the list of projects in a workspace with the Sort and Filter controls in the upper right. Click the glyph to the right of the Filter control to switch between list and card display for projects. Type a string in Search Projects… to display only projects that match the string. The project overview page appears.
In the top right corner of a monitor, click Edit Monitor.
The Edit Monitor wizard appears.
Toggle Add to Project Notifications and click Save Settings.
Notifications for this monitor are added or removed from the project according to the position of the toggle.
Operational risk
Tests for operational risk assess a model’s performance and accuracy. These tests are divided into tests for performance, drift, and abnormal input.
Performance tests
| Performance monitor | Description |
|---|---|
| Accuracy | Tests whether the model's accuracy changes relative to the reference dataset. |
| Average Thresholded Confidence (ATC) | Tests the variance of the ATC between reference and evaluation datasets. ATC estimates the accuracy of unlabeled examples. |
| Average Confidence | Tests the variance of average prediction confidence between the reference and evaluation datasets. |
| Calibration Comparison | Tests whether the calibration curve of the evaluation datset has changed relative to the reference datset. |
| Precision | Tests whether the model's precision changes relative to the reference dataset. |
| Recall | Tests whether the model's recall changes relative to the reference dataset. |
| F1 | Tests whether the model's F1 score changes relative to the reference dataset. |
| False Positive Rate | Tests whether the model's false positive rate changes relative to the reference dataset. |
| False Negative Rate | Tests whether the model's false negative rate changes relative to the reference dataset. |
| F1 | Tests whether the model's F1 score changes relative to the reference dataset. |
Drift tests
| Drift monitor | Description |
|---|---|
| Prediction Drift | Tests the change in distribution between the prediction sets generated by the reference and evaluation datasets. |
| Label Drift | Tests the change in distribution in the model's output. |
| Numeric Feature Drift | Tests the change in distribution within a given numeric feature. |
| Categorical Feature Drift | Tests the change in distribution within a given categorical feature. |
| Mutual Information Drift | Tests the change in mutual information between pairs of features, or between a feature and the label. |
| Correlation Information Drift | Tests the change in correlation between pairs of features, or between a feature and the label. |
| Embedding Drift | Tests the change in embeddings distribution between the reference and evaluation datasets. |
Abnormal inputs test
| Abnormal inputs monitor | Description |
|---|---|
| Unseen Categorical | Tests the models response to data that contains categorical values that are never observed in the reference dataset. |
| Rare Categories | Tests the models response to data that contains categorical values that are rarely observed in the reference dataset. |
| Numeric Outliers | Tests the models response to data that contains numeric values outside the typical range for that feature in the reference dataset. |
| Abnormality Rate | Tests the total percent of rows with any abnormalities. |
| Feature Type Check - Count | Tests the number of feature values that are of the incorrect type. |
| Null Check - Count | Tests the number of null values for features that do no have nulls in the reference dataset. |
| Empty String - Count | Tests the number of empty or null strings for each string feature. |
| Capitalization - Count | Tests the number of string values that are capitalized differently from those observed in the reference set. |
| Inconsistencies - Count | Tests the number of data points with pairs of feature values that are inconsistent with each other. |
| Required Characters - Count | Tests the number of required characters in feature values. |
Note: the Unseen Categorical, Rare Categories, and Numeric Outliers tests each have 2 associated monitors
Performance impact which measures the model performance change attributed to the data with the abnormality.
Count, which measures the number of occurrences of the abnormality in each bin.
Security risk
Tests for security risk assess the security of the model and underlying dataset, providing alerts in cases of model evasion or subversion.
| Security risk monitor | Description |
|---|---|
| Data Poisoning | Tests for corrupted input data. |
| Model Evasion | Tests for adversarial evasion attacks. |
Fairness and Compliance risk
Tests for fairness and compliance risk assess a model’s outcome for fair treatment among subcategories in the data.
| Fairness monitor | Description |
|---|---|
| Intersectional Group Fairness | Tests for changes in the model performance over different slices of data from the intersection of two protected features. |
| Positive Prediction Rate | Tests whether the model's positive prediction rate differs significantly across different subsets of protected features. |
| Predictive Equality | Test whether model performance differs significantly across different subsets of protected features. |
| Equal Opportunity Recall | Tests whether model recall differs significantly across different subsets of protected features. |
| Class imbalance | Test if any subsets of a feature have high class imbalance bias as a result of having a significantly smaller sample size. |
| Demographic Parity | Tests how the model performance over subsets of protected features compare to the subset with highest performance. |
| Protected Feature Drift | Tests the change in distribution of protected features. |
Events
The Overview page of a model under Continuous Testing displays a list of Active Events to the right of the active monitors.
The Events list provides several filter selectors to focus on a specific set of events.
| Filter | Description | Potential states |
|---|---|---|
| Testing | Test type | Stress Test or Continuous Test |
| Risk Categories | Major risk category | Operational, Security, Fairness |
| Status | The status of a specific test | Fail, Warning, Pass, Skip |
| Level | The importance level of the event | None, Low, High |
| Last Updated Time | The time an event was last updated | Within a time range |
Event root-cause analysis and actionability
Robust Intelligence can provide significant context and analysis of a detected event. This analysis includes the metric and threshold values defining the event and the time interval for which the metric fell below the thresholds. Events for degraded model performance monitors will show which data issue(s) may have lead to the detected event when applicable.
Sign in to a RI Platform instance.
The Workspaces page appears.
Click a workspace.
The Workspaces summary page appears.
Select a project.
You can filter or sort the list of projects in a workspace with the Sort and Filter controls in the upper right. Click the glyph to the right of the Filter control to switch between list and card display for projects. Type a string in Search Projects… to display only projects that match the string. The project overview page appears. The project overview page contains both Stress Test and Continuous Test events
(Optional) Click Show Details (only for Continuous Test events).
A description of the degradation event appears. Events with the RCA Available tag show additional context about feature drift which may have contributed to the degradation.
(Optional) Click Resolve to remove an event from the events list.
A confirmation window will apear. Click Resolve Issue to confirm.
(Optional) Click a Continuous Test event.
The CT risk page corresponding to the degraded monitor appears.