All Functions

Aggregate and metric function syntax

Functions Overview

This page provides a reference of all available functions by aggregation functions and metric functions. Click the linked model type for documentation on a particular model.

1. Aggregation functions

2. Metric functions

Aggregation functions

Every User Defined Metric must have one or more aggregation functions or metrics.

Function	Description	Type
COUNT(*)	Counts the number of rows	n/a
APPROX_COUNT_DISTINCT(exprs)	Counts the unique values of exprs	String
SUM(exprs)	Sums the value of the expression across rows	Numeric
AVG(exprs)	Averages the value of the expression across rows	Numeric
APPROX_QUANTILE(exprs,<decimal>)	Approximate quantile of expres. Second argument must be a numeric literal between 0 and 1 inclusive	Numeric
MIN(exprs)	Minimum of the value of the expression across rows	Numeric
MAX(exprs)	Maximum of the value of the expression across rows	Numeric

Metric functions

Metric functions leverage existing metrics in Arize for use in your custom metric. They also allow you to customize the way existing metrics are calculated. Metric functions can take both positional arguments and keyword arguments. When using both, positional arguments must come before keyword arguments. Keyword arguments can be specified in any order.

For classification metrics, the model's configured positive class is the default value, we will refer to this as defaultPositiveClass in this doc.

Note that these functions need actual (a.k.a. ground truth) data to produce results.

True Positive

TP(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Computes the true positive rate, using the positive class. If pos_class= is omitted, then the positive class configured for the model is used.

False Positive

FP(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Computes the false positive rate, using the positive class. If pos_class= is omitted, then the positive class configured for the model is used.

True Negative

TN(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Computes the true negative rate, using the positive class. If pos_class= is omitted, then the positive class configured for the model is used.

False Negative

FN(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Computes the false negative rate, using the positive class. If pos_class= is omitted, then the positive class configured for the model is used.

Precision

PRECISION(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Computes precision, using the positive class. If pos_class= is omitted, then the positive class configured for the model is used.

Recall

RECALL(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Computes recall, using the positive class. If pos_class= is omitted, then the positive class configured for the model is used.

F1

F1(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Compute the F1 score, also known as balanced F-score or F-measure. If pos_class= is omitted, then the positive class configured for the model is used.

F_BETA

F_BETA(actual=categoricalActualLabel, predicted=categoricalPredictionLabel, pos_class=defaultPositiveClass, beta=1)

Model Type: Score Categorical

Computes F-score with optional beta= parameter for re-weighting precision and recall. Beta is defaulted to 1, which produces the same result as the F-1 score. When beta=0, F-score equals precision, when beta goes to infinity, F-score equals recall. Commonly used values for beta= are 2, which weighs recall higher than precision, and 0.5, which weighs recall lower than precision. If pos_class= is omitted, then the positive class configured for the model is used.

LOG_LOSS

LOG_LOSS(actual=categoricalActualLabel, predicted=scorePredictionLabel, pos_class=defaultPositiveClass)

Model Type: Score Categorical

Computes log loss of the model. Note that actual= is a string column while predicted= is a numeric column.

ACCURACY

ACCURACY(actual=categoricalActualLabel, predicted=categoricalPredictionLabel)

Model Type: Score Categorical

Computes accuracy of the model.

MAE

MAE(actual=scoreActualLabel, predicted=scorePredictionLabel)

Model Type: Numeric, Score Categorical

Computes mean absolute error.

MAPE

MAPE(actual=scoreActualLabel, predicted=scorePredictionLabel)

Model Type: Numeric, Score Categorical

Computes mean absolute percentage error.

MSE

MSE(actual=scoreActualLabel, predicted=scorePredictionLabel)

Model Type: Numeric, Score Categorical

Computes mean squared error.

RMSE

RMSE(actual=scoreActualLabel, predicted=scorePredictionLabel)

Model Type: Numeric, Score Categorical

Computes root mean square error.

AUC

AUC(actual=scoreActualLabel, predicted=scorePredictionLabel)

Model Type: Score Categorical

Computes the (ROC) AUC.

NDCG

NDCG(ranking_relevance=relevance, prediction_group_id=predictionGroupId, rank=rank, omit_zero_relevance=True, k=10)

Model Type: Ranking

Computes the Normalized Discounted Cumulative Gain of a ranking model. In order to control the behavior of whether rows with 0-relevance are included or not, use omit_zero_relevance since this will impact the averaging that is implicit in this metric.