MLflow options

MLflow Model output

Use threshold from the current version of the model ({{modelDetails.userMeta.activeClassifierThreshold | number: 3}}) Override the threshold from the model

Override with

Treatment recommendation

Enable

Treatment recommendation based on treatment effect distribution is not available for multi-valued treatments

Assignment method

Treatment recommendation with exact ratio requires to load the full dataset in memory.

Treatment ratio

Defines the ratio of the population that should be treated. The treatment is recommended for samples with the highest predicted effect.

Threshold of predicted effect

The treatment is recommended for samples with a predicted effect above this threshold.

For classifications, the threshold should be in the (-1, 1) interval.

Treatment analysis

Confidence score

Minimum confidence score

Use confidence score threshold from the current version of the model ({{modelDetails.userMeta.activeClassifierThreshold | number: 2}}) Override the confidence score threshold from the model

Override with

Batch size

Output

Treatment variable

Control value

Treatment values

Preferred outcome class

Output probabilities

output probabilities for each class in addition to the prediction

Output percentile

percentile, among test set, of the proba of class 1

The input dataset must contain the past data at least {{ prettyTimeSteps(modelDetails.iperf.minTimeseriesSizeForScoring * modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }} of past data (up to {{ prettyTimeSteps(modelDetails.iperf.maxUsedTimestepsForScoring * modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }} will be used for scoring) and {{ prettyTimeSteps((desc.predictionLength || 0) * modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }} of future external features values ({{ 'column' | plurify: externalFeatures().length }} {{ externalFeatures().join(', ') }}.)

Forecast length

Number of steps to forecast
(must be smaller or equal to horizon during training: {{ prettyTimeSteps(modelDetails.coreParams.predictionLength * modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }}). (horizon during training is {{ prettyTimeSteps(modelDetails.coreParams.predictionLength * modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }}).

The requested forecast maximum timestamp might exceed the maximum valid timestamp 2262-04-11.
The scoring recipe might fail.

The requested forecast length exceeds the maximum index available (approximately 584 years, or 2^64 nanoseconds).
The scoring recipe will fail.

Recursive scoring not available for this algorithm.
Forecast length must be smaller or equal to the size of the training horizon: {{ modelDetails.coreParams.predictionLength }}.

Recursive scoring will be performed and quantiles estimations won't be available for forecast times beyond {{ prettyTimeSteps(modelDetails.coreParams.predictionLength * modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }}.

Forecast quantiles

Output quantiles for every forecast time step

Not supported by algorithm.

Past data to include

Number of past time steps to include in the output (-1 for all).

Refit model

Refit model on input data (only relevant for statistical models, mandatory for Seasonal trend).

Output explanations

Compute individual explanations (much slower)

Computation method

ICE Shapley

(more robust but slower)

Number of explanations

Most influential explanations to compute

Sub chunk size

Higher means faster but larger memory usage

Number of Monte Carlo steps

Higher means more robust but slower computation (between 25 and 1000)

Use input as explanation basis

Draw the random sample from the input dataset instead of the model's train set

Model parameters

Forecast horizon during training

{{ prettyTimeSteps(modelDetails.coreParams.predictionLength * modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }} ({{ modelDetails.coreParams.predictionLength }} {{ 'step' | plurify: modelDetails.coreParams.predictionLength }} of {{ prettyTimeSteps(modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }}).

Time step parameters

Time step

{{ prettyTimeSteps(modelDetails.coreParams.timestepParams.numberOfTimeunits, modelDetails.coreParams.timestepParams.timeunit) }}

Alignment

{{ prettySelectedDate(modelDetails.coreParams.timestepParams.timeunit, modelDetails.coreParams.timestepParams.monthlyAlignment, modelDetails.coreParams.timestepParams.unitAlignment) }}

{{ getWeekDayName(modelDetails.coreParams.timestepParams.endOfWeekDay) }}

Resampling parameters

Numerical interpolation

Numerical extrapolation

Non-numerical imputation

Duplicate

Engine

Warning : Conditional columns will not be output with in-database scoring.

Warning : the algorithm you are scoring may create a very large SQL query, resulting in slow query computation and run time and possibly query failure.

Note : the algorithm you are scoring does not support probability outputs with a SQL engine. To obtain class probabilities, select another engine.

Warning : Calibration is not supported with SQL engine, output will be computed without taking calibration into account. To obtain class probabilities with calibration, select another engine.

No settings are available for regression scoring