AITransform for Jedox Cloud

Time Series

Choose "TimeSeriesForecasting" in the Function field. This will allow for simple time series forecasting methods with a basic interpolation option. If you would like to have more control over interpolation and outlier methods in your time series prediction, please choose the AIssistedTimeSeriesPipeline function.

Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

The name of the measures dimension to store the prediction values.

Input source	Input source is an extract or transform containing at least Time, Version, and Measures dimensions. Additional dimensions should also be included if they are important to your data structure, so for example, when using a cube extract. The input data should be time series data you would like to base your prediction on, with as few missing values as possible. The time dimension should be structured as either daily, weekly, monthly, or yearly data. The data should be normalized. An example source input cube extract of monthly data for two measures is shown below. Notice the time query that ensures the extract pulls only data monthly data and not, for example a time element like 2020-00. Note: do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time, Version, and Measures dimensions can have variable or constant inputs.
Values	The Values column from the time series input source, e.g. #Value.
Function	Choose Time Series Forecasting from the list of AI functions. It uses historic data to generate forecasts based on past trends.
Period	Input data must have an assigned period, which is defined as a number of data points per period. Period can be Daily (365), Monthly (12), Weekly (52), and Yearly (1). In the example shown above, it would be Monthly.
Output source	Target extract/transform where the results values are stored. The output source must already exist before the transform is run. The dimension mapping must be similar to that of the Input Source, i.e., containing Measures, Time, and Version dimensions. However, the structure might change, depending on the output type you select in the Parameters. Please see Output Type in the Parameters section below. Below is an example of an output source cube extract that belongs to the Input Source example above. It looks similar to the input, however there are some differences. Output Type 2 is selected in the parameters, which means the Version dimension should have three elements in which to populate the values of predictions, upper bound predictions, and lower bound predictions. The time dimension filter is also different, as the prediction values should be populated into the year in which they belong, in this case 2021. From here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data, or export it to another data source. Note: do not use the input source as an output source, as the input data may be replaced if the transform is followed by a cube load.
Output Time dimension	The name of the time dimension where you would like result values stored.
Output Version dimension	The name of the version dimension where you would like the result values stored.
Output Measures dimension	The name of the measures dimension where you would like the result values stored.
Output values	The name of the values column where your result values will be populated, e.g. #Values.

Parameters - Time Series

The Function parameters are preset by the choices you made in the above Function setting.

Function Parameter	Description
NumberOfPredictions	Sets the number of time units (as set in Period parameter) to predict, i.e. 5 years, 12 months, 180 days, etc. It is preset for monthly input; you can change this in the Period parameter in Settings.
Algorithm Type	The AITransform includes 12 algorithms for data analysis: 1 - Linear Model 2 - Holt Winter 3 - Seasonal Naïve 4 - Exponential Smoothing 5 - ARIMA 6 - Random Walk with Drift 7 - Seasonal and Trend Decomposition using Loess (STL) model 8 - generalized STL model 9 - Neural Network Time Series Forecast 10 - TBATs model 11 - Croston's Method 12 - Extrapolation (note that extrapolation is not AI or Machine Learning) If "all" or "0" is entered, then all algorithms will run and the most accurate will be chosen for result. Enter 1-12 to select a particular algorithm. Multiple algorithms can be entered with comma separation, e.g. "1,3,5". The most accurate of the selected algorithms will be chosen for the result.
Output type	0 - Output consists of rows that correspond to NumberOfPredictions selected, e.g. 12 months results in output extract with 12 rows of prediction values. 1 - Output consists of columns for each of the algorithms selected, e.g. 12 algorithms result in output extract with 12 columns of training accuracy values. 2 - NumberOfPredictions as in output type 0, plus 2 more sets of rows with lower bounds and upper bounds. For example, if NumberOfPredictions is set for 3 months, then the resulting extract will consist of 3 rows of prediction output, plus 3 rows of lower bounds and 3 rows of upper bounds. 3 - Everything from above: NumberOfPredictions rows, lower bounds, upper bounds, as well as extra columns for each of the algorithms selected (12 algorithms = 12 columns). In the final project output, these columns could be separated/removed with a FieldTransform.
Accuracy threshold	Minimum percentage of training accuracy required from the algorithms in order to return values, e.g. 40 = 40% accuracy must be met or no values will be returned.
UseInterpolation	true: zero or null values in the source data will be replaced with meaningful values before predictions are calculated. false: zero or null values will be included in the source data sent to the algorithms.
PercentageOfZeros	Percentage of zero/null data that will be tolerated. When the threshold is reached, an error message appears ("one or more time series has more than x% zeros").

Interpolation

This function replaces missing values in source data with meaningful values for better results. Choose "Interpolation" in the Function field. Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract/transform containing Time, Version, and Measures dimensions. Additional dimensions should also be included if they are important to your data structure, so for example, when using a cube extract. The input data should be the entire time series data you would like to have analyzed to find missing values and have a replacement suggested. The time dimension should be structured as either daily, weekly, monthly, or yearly data. The data should be normalized. An example source input cube extract of monthly data for one measure is shown below. Notice the time query that ensures the extract pulls only data monthly data and not, for example a time element like 2020-00. From here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Note: do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time, Version, and Measures dimensions can have variable or constant inputs.
Values	The Values column from the time series input source, e.g. #Value.
Function	Interpolation: replaces missing values in source data with meaningful values for better results.
Period	Input data must have an assigned period, which is defined as a number of data points per period. Period can be Daily (365), Monthly (12), Weekly (52), and Yearly(1). In the example shown above, it would be Monthly.
Output source	Target extract/transform where the results values are stored. The output source must already exist before the transform is run. The dimension mapping must be similar to that of the Input Source, i.e., containing Measures, Time, and Version dimensions. The only change should be in the Version element name where the interpolated values should be stored, e.g. Interpolation instead of Actual. Below is an example of an output source cube extract that belongs to the Input Source example above. Note: do not use the input source as an output source, as the input data may be replaced if the transform is followed by a cube load. Also, Input Source and Output Source must have the same number of rows.
Output Time dimension	The name of the time dimension where you would like result values stored.
Output Version dimension	The name of the version dimension where you would like the result values stored.
Output Measures dimension	The name of the measures dimension where you would like the result values stored.
Output values	The name of the values column where your result values will be populated.

Parameters - Interpolation

The Function parameters are preset by the choices you made in the above setting. If you would like to make changes, however, here are the possible inputs.

AlgGroup

Interpolation algorithms are divided into three groups as mentioned below:

TABLE 1: Group Types for Interpolation service
Group	Value
Simple	0
Seasonal	1
Advanced	2

AlgorithmType and AlgSubType

Each algorithm group is further divided into different algorithms and their subtypes (AlgSubType) as mentioned in the table below:

TABLE 2: Group, Algorithm, and Subtype combinations for Interpolation
Group	Algorithm	Subtype	Description
0	0	Not applicable	Last observation carried forward (LOCF).
	1	Not applicable	Linear approximation.
	2	Not applicable	Spline.
	3	Not applicable	Random sampling.
	4	0	Moving average with equal weights for all observations.
		1	Moving average with linear weights.
		2	Moving average with exponential weights.
	5	0	Mean
		1	Mode
		2	Median
1	0	0	Seasonal decomposition with LOCF.
		1	Seasonal decomposition with mean.
		2	Seasonal decomposition with linear approximation.
	1	0	Seasonal split with LOCF.
		1	Seasonal split with mean.
		2	Seasonal split with linear approximation.
2	0	0	Kalman smoothing with basic ARIMA model.
	0	1	Kalman smoothing with basic state space model.
	1	Not applicable	k-nearest neighbors with median.
	2	Not applicable	Linear regression model.

This parameter is currently not used in the processing (it is recommended to use the default setting for now), but works as follows:

0	Interpolated values (default setting).
1	Positions of missing values (The positions are identified as 1 for 0 in input, 2 for NA in input and 0 otherwise.)
2	Interpolated values and positions.

OutlierTS

This function detects extreme values in source data and replaces them with meaningful values for more accurate forecasts.

Choose "OutlierTS" in the Function field. Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract/transform containing Time, Version, and Measures dimensions. Additional dimensions should also be included if they are important to your data structure, so for example, when using a cube extract. The input data should be the entire time series data you would like to have analyzed to find missing values and have a replacement suggested. The time dimension should be structured as either daily, weekly, monthly, or yearly data. The data should be normalized. An example source input cube extract of monthly data for one measure is shown below. Notice the time query that ensures the extract pulls only data monthly data and not, for example a time element like 2020-00. Note: do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time, Version, and Measures dimensions can have variable or constant inputs.
Values	The Values column from the time series input source, e.g. #Value.
Function	Outlier Detection: detects extreme values in source data and replaces them with meaningful values for more accurate forecasts.
Period	Input data must have an assigned period, which is defined as a number of data points per period. Period can be Daily (365), Monthly (12), Weekly (52), and Yearly(1).
Output source	Target extract/transform where the results values are stored. The output source must already exist before the transform is run. The dimension mapping must be similar to that of the Input Source, i.e., containing Measures, Time, and Version dimensions. The only change should be in the Version element name where the interpolated values should be stored, e.g. Outlier instead of Actual. Below is an example of an output source cube extract that belongs to the Input Source example above. From here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Note: do not use the input source as an output source, as the input data may be replaced if the transform is followed by a cube load. Also, Input Source and Output Source must have the same number of rows.
Output Time dimension	The name of the time dimension where you would like result values stored.
Output Version dimension	The name of the version dimension where you would like the result values stored.
Output Measures dimension	The name of the measures dimension where you would like the result values stored.
Output values	The name of the values column where your result values will be populated.

Parameters - OutlierTS

The Function parameters are preset by the choices you made in the above setting.

AlgorithmType

Algorithms available for Outlier detection:

Algorithm	Description
IO	Innovational outlier detection.
AO	Additive outlier detection.
LS	To detect level shifts in the time series.
TC	To detect temporary changes.
SLS	To consider seasonal level shifts while detecting outliers.
TSO	Basic time series outlier detection.
CD	Cooks’s distance for outlier detection using linear modeling.
EV	Extreme value detection.

UseInterpolation

Generate replacement values for detected outliers (recommended to use 0 or 1).

0	Default replacement using linear approximation.
1	User-defined interpolation for replacement (AlgGroup, AlgInterpolate, AlgSubType parameters are used).
2	No replacement.

AlgGroup

Same as AlgGroup parameter in Interpolation.

AlgInterpolate

Same as AlgorithmType in Interpolation.

AlgSubType

Same as AlgSubType in Interpolation.

OutputType

This parameter is currently not used in the processing, (it is recommended to use the default setting for now) but works as follows:

0	Replacement values for detected outliers.
1	Positions of outlying values (the positions are identified as 1 for outlier in input and 0 otherwise).
2	Replacement values for detected outliers and positions.
3	Replacement values for detected outliers and positions. 0 is not considered as an outlier when this is used.

FailSafe

Outlier detection algorithms are highly data-dependent, so all algorithms may not work for all kinds of data. This option decides the fall back in case the algorithm selected by the user fails.

0	No fail-safe.
1	EV as a fail-safe algorithm.
2	Run all algorithms until the first valid output is produced.

AIssistedDataPreparation

This function offers suggestions for missing values and identifies/replaces outliers based on user input to calculate results.

Choose "AIssistedDataPreparation" in the Function field. Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract/transform containing Time, Version, and Measures dimensions. Additional dimensions should also be included if they are important to your data structure, so for example, when using a cube extract. The input data should be the entire time series data you would like to have analyzed. The time dimension should be structured as either daily, weekly, monthly, or yearly data. The data should be normalized. An example source input cube extract of monthly data for one measure is shown below. Notice the time query that ensures the extract pulls only data monthly data and not, for example a time element like 2020-00. Note: do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time, Version, and Measures dimensions can have variable or constant inputs.
Values	The Values column from the time series input source, e.g. #Value.
Function	AIssistedDataPreparation: manipulates missing values and identifies/replaces outliers based on user input to calculate results.
Period	Input data must have an assigned period, which is defined as a number of data points per period. Period can be Daily (365), Monthly (12), Weekly (52), and Yearly(1).
Output source	The dimension mapping must be similar to that of the Input Source, i.e., containing Measures, Time, and Version dimensions. However, the structure might change, depending on the output type you select in the Parameters. Please see Output Type in the Parameters section below. Below is an example of an output source cube extract that belongs to the Input Source example above. It looks similar to the input, however there are some differences. Output Type “all” was selected in the parameters, which means the Version dimension should have three elements in which to populate the values of Outlier, Interpolation and Extrapolation. Depending on the output type, the output source will change. From here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Note: do not use the input source as an output source, as the input data may be replaced if the transform is followed by a cube load.
Output Time dimension	The name of the time dimension where you would like result values stored.
Output Version dimension	The name of the version dimension where you would like the result values stored.
Output Measures dimension	The name of the measures dimension where you would like the result values stored.
Output values	The name of the values column where your result values will be populated.

Parameters - AIssistedDataPreparation

The Function parameters are preset by the choices you made in the above setting.

Tasks

Defines the data preparation task to be performed.

all	All tasks are performed individually.
oie	All tasks are performed sequentially.
oi	Outlier detection and interpolation.
oe	Outlier detection and extrapolation.
io	Interpolation and outlier detection.
ie	Interpolation and extrapolation.
ei	Extrapolation and interpolation.
eo	Extrapolation and outlier detection.
o	Outlier.
i	Interpolation.
e	Extrapolation.

Notes

Oie, oi, oe, io, ie, ei, eo are performed sequentially (the output of one method is used as input for the next).

If the task value is "all", then three times as many rows are needed for the output as the input has. The output needs to include the three result version element names, which are "Interpolation", "Outlier", and "Extrapolation".

For all other task values (e.g. oie, oi, ei, i, etc.), the output can have the same number of rows as the input, and the output should only have one version element name.

AlgorithmOutlier

The algorithm to be used for outlier detection. For details see the OutlierTS service.

FailSafe

Outlier detection algorithms are highly data-dependent, so all algorithms may not work for all kinds of data. This option decides the fall back in case the algorithm selected by the user fails.

0	No fail-safe.
1	EV as a fail-safe algorithm.
2	Run all algorithms until the first valid output is produced.

AlgorithmInterpolate

Same as AlgorithmType in Interpolation.

AlgorithmExtrapolate

The algorithm to be used for extrapolation:

per	periodic spline
lin	linear spline
cub	cubic spline
tsf	Time Series forecasting (best prediction from algorithms 1 to 11)
tsf1-tsf11	Time Series forecasting (algorithms 1 to 11)

AIssistedTimeSeriesPipeline

Choose "AIssistedTimeSeriesPipeline" in the Function field. This function allows more control over interpolation and outlier methods in your time series prediction. For simple time series forecasting methods with a basic interpolation option, please choose the TimeSeries function.

Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract or transform containing at least Time, Version, and Measures dimensions. Additional dimensions should also be included if they are important to your data structure, so for example, when using a cube extract. The input data should be time series data you would like to base your prediction on, with as few missing values as possible, although the service will forecast using interpolated values. The time dimension should be structured as either daily, weekly, monthly, or yearly data. The data should be normalized. An example source input cube extract of monthly data for two measures is shown below. Notice the time query that ensures the extract pulls only data monthly data and not, for example a time element like 2020-00. Note: do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time, Version, and Measures dimensions can have variable or constant inputs.
Values	The Values column from the time series input source, e.g. #Value.
Function	AIssistedTimeSeriesPipelineInput: uses outlier detection, interpolation, and extrapolation services based on user input to prepare data and then forecast values.
Period	Input data must have an assigned period, which is defined as a number of data points per period. Period can be Daily (365), Monthly (12), Weekly (52), and Yearly(1). In the example shown above, it would be Monthly.
Output source	Target extract/transform where the results values are stored. The output source must already exist before the transform is run. The dimension mapping must be similar to that of the Input Source, i.e., containing Measures, Time, and Version dimensions. Below is an example of an output source cube extract that belongs to the Input Source example above. It looks similar to the input to the input, however there are some differences. The Version dimension should have three elements in which to populate the values of predictions, upper bound predictions and lower bound predictions. The time dimension filter is also different, as the prediction values should be populated into the year in which they belong, in this case 2022. From here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Note: do not use the input source as an output source, as the input data may be replaced if the transform is followed by a cube load.
Output Time dimension	The name of the time dimension where you would like result values stored.
Output Version dimension	The name of the version dimension where you would like the result values stored.
Output Measures dimension	The name of the measures dimension where you would like the result values stored.
Output values	The name of the values column where your result values will be populated, e.g. #Values.

Parameters - AIssistedTimeSeriesPipeline

The Function parameters are preset by the choices you made in the above setting.

NumberOfPredictions

Sets the number of time units (as set in Period parameter) to predict, i.e. 5 years, 12 months, 180 days, etc. It is preset for monthly input; you can change this in the Period parameter in Settings.

AlgorithmType

The AITransform includes 12 algorithms for data analysis:

1 - Linear Model
2 - Holt Winter
3 - Seasonal Naïve
4 - Exponential Smoothing
5 - ARIMA
6 - Random Walk with Drift
7 - Seasonal and Trend Decomposition using Loess (STL) model
8 - generalized STL model
9 - Neural Network Time Series Forecast
10 - TBATs model
11 - Croston's Method
12 - Extrapolation (note that extrapolation is not AI or Machine Learning)

If "all" or "0" is entered, then all algorithms will run and the most accurate will be chosen for result.

Enter 1-12 to select a particular algorithm. Multiple algorithms can be entered with comma separation, e.g. "1,3,5". The most accurate of the selected algorithms will be chosen for the result.

AccuracyThreshold

Minimum percentage of training accuracy required from the algorithms in order to return values, e.g. 40 = 40% accuracy must be met or no values will be returned.

PercentageOfZeros

Percentage of zero/null data that will be tolerated. When the threshold is reached, an error message appears ("one or more time series has more than x% zeros").

NumberOfTestValues

The number of rows to be taken as test set for each time series.

DataPreparationTask

Defines the data preparation task to be performed.

all	All tasks are performed individually.
oie	All tasks are performed sequentially.
oi	Outlier detection and interpolation.
oe	Outlier detection and extrapolation.
io	Interpolation and outlier detection.
ie	Interpolation and extrapolation.
ei	Extrapolation and interpolation.
eo	Extrapolation and outlier detection.
o	Outlier.
i	Interpolation.
e	Extrapolation.

Note: oie, oi, oe, io, ie, ei, eo are performed sequentially (the output of one method is used as input for the next).

AlgorithmOutlier

The algorithm to be used for outlier detection. For details see the OutlierTS service.

FailSafe

Outlier detection algorithms are highly data-dependent, so all algorithms may not work for all kinds of data. This option decides the fall back in case the algorithm selected by the user fails.

0	No fail-safe.
1	EV as a fail-safe algorithm.
2	Run all algorithms until the first valid output is produced.

AlgorithmInterpolate

Same as AlgorithmType in Interpolation.

AlgorithmExtrapolate

The algorithm to be used for extrapolation:

per	periodic spline
lin	linear spline
cub	cubic spline
tsf	Time Series forecasting (best prediction from algorithms 1 to 11)
tsf1-tsf11	Time Series forecasting (algorithms 1 to 11)

DriverBasedPrediction

This function performs basic pre-processing (if required) on input data and trains a regression model for predicting target values. It is often the second type of prediction used if time series prediction doesn’t produce accurate results.

Choose "DriverBasedPrediction" in the Function field. Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract or transform containing at least Time, Version, and Measures dimensions. Additional dimensions should also be included if they are important to your data structure, so for example, when using a cube extract. The input data should be data you would like to base your prediction on, with as few missing values as possible. The time dimension should be structured as either daily, weekly, monthly, or yearly data. The data should be normalized. An example source input cube extract of daily data for several measures is shown below. All drivers, including the target measure you want to predict should be in this source data. If you are unsure which drivers to choose, try using DriverSelection. Notice that other dimensions are listed in the query filters as well, most notably the time dimension queries. In this example, this ensures that the source data is passed as daily data, and not quarterly. Note: do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time, Version, and Measures dimensions can have variable or constant inputs.
Values	The Values column from the time series input source, e.g. #Value.
Function	DriverBasedPrediction: performs basic pre-processing (if required) on input data and trains a regression model for predicting target values.
Period	Input data must have an assigned period, which is defined as a number of data points per period. Period can be Daily (365), Monthly (12), Weekly (52), and Yearly(1). The example above is daily.
Output source	Target extract/transform where the results values are stored. The output source must already exist before the transform is run. The dimension mapping must be similar to that of the Input Source, i.e., containing Measures, Time, and Version dimensions. Below is an example of an output source cube extract that belongs to the Input Source example above. All drivers, including the target measure you want to predict should also be included this output source data. Note: In order to get meaningful predictions for your target measure, you must have values for all other drivers in this output source. Without these values, the prediction will likely be useless. From here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Aside from the prediction values, the transform will also give accuracy values for each algorithm used. See the Data Preview example below. Note: do not use the input source as an output source, as the input data may be replaced if the transform is followed by a cube load.
Output Time dimension	The name of the time dimension where you would like result values stored.
Output Version dimension	The name of the version dimension where you would like the result values stored.
Output Measures dimension	The name of the measures dimension where you would like the result values stored.
Output values	The name of the values column where your result values will be populated, e.g. #Value.

Parameters - DriverBasedPrediction

The Function parameters are preset by the choices you made in the above setting.

Function Parameter

Description

AlgorithmType

Best Selection	0
Logistic Regression	1
Support Vector Machine	2
Random Forest	3
Extra Trees	4
AdaBoost Regressor	5
Gradient Boosting Regressor	6
KNN Regression	7

OptimizeParameters

0

Normalize

False

FixRandomState

True

TargetMeasure

Measure to be predicted / dependent measure. In this example, Units.

TargetVersion

Name of element in the Version dimension in which to store the predicted values. In this example, Prediction.

DriverSelection

This function performs basic pre-processing (if required) on input data and identifies the important input drivers with respect to a given target.

Choose "DriverSelection" in the Function field. Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract/transform containing Time, Version, and Measures dimensions. Additional dimensions should also be included if they are important to your data structure, so for example, when using a cube extract. An example source input cube extract of daily data for all measures is shown below. Notice the time query that ensures that the extract pulls only complete data, in this case up until the end of 2021 and does not include quarterly elements.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time, Version, and Measures dimensions can have variable or constant inputs.
Values	The Values column from the time series input source, e.g. #Value.
Function	DriverSelection: performs basic pre-processing (if required) on input data and identifies the important input features with respect to a given target.
Period	Input data must have an assigned period, which is defined as a number of data points per period. Period can be Daily (365), Monthly (12), Weekly (52), and Yearly(1). The above example would be Daily.

Parameters - DriverSelection

The Function parameters are preset by the choices you made in the above setting.

Function Parameter

Description

AlgorithmType

Random Forest based Importance detection	1, 2
Generalized Linear Model based significance calculation	3

Normalize

False

FixRandomState

True

TargetMeasure

Measure to be analyzed / dependent measure, in this example, Units.

Because there is no output source, from here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Depending on the algorithm type, the output will be different. Below is an example of a Data Preview with the AlgorithmType 3.

This shows the significance level of each feature (the higher the number, the greater the significance of the driver) and the importance values (the closer the driver’s value is to 0, the more important it is).

Classification

This function performs basic pre-processing (if required) on input data and trains a classifier for predicting class labels. Choose "Classification" in the Function field. Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract/transform containing Time (or unique identifier) and Version dimensions. The input for Classification function must be denormalized, so a cube extract on its own will not work. The input source is a denormalized set of data that includes all dimensions and measures that are important to the classification data set as well as values for the target, the class you want to predict, e.g. Churn. Note: do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time and Version can have variable or constant inputs. The Time dimension can also be a unique identifier of each row as well, for example CustomerID.
Function	Classification: performs basic pre-processing (if required) on input data and trains a classifier for predicting class labels.
Output source	Target extract/transform where the results values are stored. The output source must already exist before the transform is run. The dimension mapping must be similar to that of the Input Source, i.e., containing Time/Unique ID and Version dimensions. The output source for Classification function must also be denormalized. From here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Aside from the prediction values, the transform will also give accuracy values for each algorithm used. The Data Preview example below shows the churn class prediction (1.0 or 2.0) as well as the accuracy for a few algorithms used. Not shown are the other dimensions (CustomerID and others) used in the input/output source. Do not use the input source as an output source, as the input data may be replaced if the transform is followed by a cube load.
Output Time dimension	The name of the time or unique identifier dimension where you would like result values stored.
Output Version dimension	The name of the version dimension where you would like the result values stored.

Parameters - Classification

The Function parameters are preset by the choices you made in the above setting.

Function Parameter

Description

AlgorithmType

Best Selection, which runs all algorithms and returns the values from the algorithm that performed best	0
Logistic Regression	1
Support Vector Machine	2
Extra Trees	3
Random Forest	4
Decision Trees	5
GaussianNB	6
MultinomialNB	7
ComplementNB	8

Normalize

False

FixRandomState

True

ListOfMeasures

A list of all measures used in the data set, comma delimited.

TargetVersion

Element name in Version dimension in which you would like results populated, e.g. Prediction

TargetMeasure

Measure to be analyzed / dependent measure, e.g. Churn.

DriverSelectionClassification

This function can solve classification problems where the goal is to predict the category or class that a data point belongs to, instead of predicting a numerical value, as is the case with standard DriverSelection.

Choose "DriverSelectionClassification" in the Function field. Starting with the function field will ensure you see the appropriate fields and parameters to select for each respective function.

Settings

Input source	Input source is an extract/transform containing Time (or unique identifier) and Version dimensions. The input for Classification function must be denormalized, so a cube extract on its own will not work. The input source is a denormalized set of data that includes all dimensions and measures that are important to the classification data set as well as values for the target, the class you want to predict, e.g. Churn. Note: Do not use the same source for both input and export, as the input data may be replaced if the transform is followed by a cube load.
Time dimension Version dimension Measures dimension	The dropdown lists contain column names from the input source. Time and Version can have variable or constant inputs. The Time dimension can also be a unique identifier of each row as well, for example CustomerID. The Measures dimension should be disabled for Driver Selection Classification. If it is not, ignore it.
Values	This option should be disabled for Driver Selection Classification. If it is not, ignore it.
Function	DriverSelectionClassification: performs basic pre-processing (if required) on input data and identifies the important input features with respect to a given target. This function can solve classification problems where the goal is to predict the category or class that a data point belongs to, instead of predicting a numerical value, as is the case with standard DriverSelection.
Period	This option should be disabled for Driver Selection Classification. If it is not, ignore it.

Parameters - DriverSelectionClassification

The Function parameters are preset by the choices you made in the above setting.

Function Parameter

Description

AlgorithmType

Random Forest based Importance detection	1, 2
Generalized Linear Model based significance calculation	3

Normalize

False

FixRandomState

True

TargetMeasure

Measure to be analyzed / dependent measure, Measure to be analyzed / dependent measure, e.g. Churn.

ListOfMeasures

A list of all measures used in the data set, comma delimited.

Because there is no output source, from here you can transform your data as you would like in order to, for example, load it to a cube, join it with other data or export it to another data source. Depending on the algorithm type, the output will be different. Below is an example of a Data Preview with the AlgorithmType 3.

This shows the significance level of each feature (the higher the number, the greater the significance of the driver) and the importance values (the closer the driver’s value is to 0, the more important it is).

AITransform for Jedox Cloud

AITransform functions

Settings

Parameters - Time Series

Settings

Parameters - Interpolation

AlgGroup

AlgorithmType and AlgSubType

Settings

Parameters - OutlierTS

Settings

Parameters - AIssistedDataPreparation

Settings

Parameters - AIssistedTimeSeriesPipeline

Settings

Parameters - DriverBasedPrediction

Settings

Parameters - DriverSelection

Settings

Parameters - Classification

Settings

Parameters - DriverSelectionClassification