Classification evaluation
The Classification evaluation view returns performance metrics for a deployed classification model.
Use it to evaluate binary classification performance, inspect threshold-dependent trade-offs, and review global metrics such as PR AUC and ROC AUC.
When to use
Generate the view
Generate the Classification evaluation view from a deployed model.
The request needs view_type set to CLASSIFICATION_EVALUATION.
You can evaluate either a submitted query batch with target observations, or a precomputed observation split using observation_type.
When evaluating query data, provide both query and targets. When evaluating precomputed observations, set observation_type to TRAIN, TEST, or ALL.
Output
The response follows the shared Response structure format.
The Classification evaluation view returns multiple named dataframes inside view_data.
Interpret the result
The Classification evaluation view is easiest to read in two passes: use global metrics to understand overall separability, then use threshold metrics to choose or review an operating point.
Start with global metrics
Use global_metrics to understand the model’s overall classification performance.
roc_auc measures how well the model separates classes across thresholds. pr_auc summarizes precision-recall performance and is especially useful when the positive class is rare or class balance matters.
Global metrics are useful for comparing models or splits, but they do not tell you which decision threshold to use.
Review threshold metrics
Use threshold_metrics to understand performance at each decision_threshold.
Each row shows how the model behaves when predicted probabilities are converted into binary predictions at that threshold. Review the confusion counts first, then use derived metrics such as precision, recall, F1 score, accuracy, and MCC.
Choose an operating point
Use decision_threshold to compare the trade-off between false positives and false negatives.
Lower thresholds usually increase positive predictions, which may increase recall but also increase false positives. Higher thresholds usually reduce positive predictions, which may improve precision but increase false negatives.
Choose the threshold that best matches the cost of each error type in your application.
Inspect confusion counts
Use true_positives, false_negatives, false_positives, and true_negatives to understand the raw classification outcomes at each threshold.
These counts are often easier to reason about than derived metrics, especially when communicating performance to non-technical stakeholders.
Compare evaluation sources
Use observation_type when you want to evaluate train, test, or all precomputed observations.
Comparing TRAIN and TEST results can help identify overfitting or performance drift between the data used to train the model and the data used to evaluate it.
Options
The Classification evaluation view supports options for selecting evaluation data and thresholds.
Observation type
Use observation_type when you want the evaluation to run against precomputed observations.
When observation_type is not provided, the view evaluates the query and target data supplied in the request.
Decision threshold
Use decision_threshold to control the probability threshold used to convert predicted probabilities into class predictions.
You can provide a single number or a list of numbers between 0 and 1. The view may include additional threshold rows beyond the thresholds you provide.
Field reference
The Classification evaluation view returns dataframes, so individual columns are not documented as standalone API schema properties.
threshold_metrics
The threshold_metrics dataframe contains metrics calculated at each decision threshold.
Log loss is the negative log-likelihood of the predicted probabilities. It penalizes confidently incorrect predictions.
global_metrics
The global_metrics dataframe contains threshold-independent metrics.