Test the Key Influencer service

In order to ease the readability of this tutorial, the C4PAURL token was used
to replace the predictive services Application URL displayed on the overview page.

Therefore you can replace any occurrence of the token by your value listed.

The Application URL should look like this (where XYZ is your SAP Cloud Platform account name):

https://aac4paservicesXYZ.hanatrial.ondemand.com/com.sap.aa.c4pa.services

If you are unclear with what is your SAP Cloud Platform account name, you can refer to the following blog entry: SAP Cloud Platform login, user name, account id, name or display name: you are lost? Not anymore!

The Key Influencers service analyzes a dataset to identify the variables with an influence on a specified target variable.

This service:

• Identifies the variables with an influence on a specified target ordered by decreasing contribution
• Returns detailed information on the grouped categories for each contributive variable
• Provides indicators on the reliability of the results

Note: The target variable must be either binary nominal or continuous. Multinomial targets are not supported.
To summarize, in order to execute the key influencer service, you will need a dataset with:

• a target variable for which you want to find the influencers variables
• a set of variables that potentially influence the target variable.

Optionally, you can define the following parameters to enhance your analysis:

• auto-selection: flag that indicates if you want to shorten the number of retained influencers
• number of influencers: the number of key influencers to return
• skipped variables: a list of variables to skip from the analysis
• target value: in the case of binary target variable, it indicates the target variable value of interest. By default the value with the lowest frequency will be picked.
• variable description: a more details description of the dataset
• weight variable: a column to be used to increase the importance of a row

The dataset will be using during this tutorial is extracted from the sample dataset available with SAP Predictive Analytics.

The Census sample data file that you will use to follow the scenarios for Regression/Classification and Segmentation/Clustering is an excerpt from the American Census Bureau database, completed in 1994 by Barry Becker.

Note: For more information about the American Census Bureau, see http://www.census.gov published on a non-SAP site.
This file presents the data on 48,842 individual Americans, of at least 17 years of age. Each individual is characterized by 15 data items. These data, or variables, are described in the following table.

First you need to register the Census dataset.

As described in Step 1 of Test the Dataset service tutorial, register the Census dataset using the following elements:

Open a new tab in Postman.

If you don’t have Postman installed yet, you can refer to the following how-to guide: Install Postman extension for Google Chrome as a REST client

Fill in the following information:

Select the Body tab, enable the raw mode and select JSON (application/json) in the drop down, then add the following content:

{
  "location": {
    "schema" : "PSDEMO",
    "table" : "Census"
  }
}

Take note of the returned dataset identifier.

Open a new tab in Postman.

Fill in the following information:

Select the Authorization tab and fill in the following information:

Note:
Your SAP Cloud Platform Account login is usually the email address used to register your SAP Cloud Platform account.

Select the Body tab, enable the raw mode and select JSON (application/json) in the drop down, then add the following content:

{
  "datasetID": 9999999,
  "targetColumn": "class",
  "numberOfInfluencers" : 2,
  "targetKey" : 1,
  "skippedVariables" : ["id", "sex", "race"]
}

Make sure the datasetID (here the value 9999999) is correct. To get the list of valid identifier, you can run Step 6: List all registered datasets from the Test the Data Set SAP Predictive services using a REST client
tutorial.
With these settings, you will get the top 2 influencers of the class variable, excluding the id, sex, race variables from the analysis.

Click on Send

Congratulations! You have just run a key influencer service on the Census dataset.

The top 2 influencers are the marital_status and capital_gain.

They have been picked because they have the highest contribution in the model. The model also provide you the groups resulting from the encoding phase.

But if you look closer you will notice that the capital_gain has been considered as nominal instead of continuous.

Let’s fix that with a more accurate variable description.

Provide an answer to the question below then click on Validate.

Just like in the Forecast service, you could adjust the variable description at the dataset level, but you will instead use the Key Influencers service parameters to adjust the variable description.

Open a new tab in Postman.

Fill in the following information

Select the Authorization tab and fill in the same details as in the previous call.

Select the Body tab, enable the raw mode and select JSON (application/json) in the drop down, then add the following content:

{
  "datasetID": 9999999,
  "targetColumn": "class",
  "numberOfInfluencers" : 2,
  "targetKey" : 1,
  "skippedVariables" : ["id", "sex", "race"],
  "variableDescription" : [
    {"position" : "1", "variable" : "id", "storage" : "integer" , "value" : "nominal" ,  "key" : "1"},
  	{"position" : "2", "variable" : "age", "storage" : "integer" , "value" : "continuous"},
  	{"position" : "3", "variable" : "workclass", "storage" : "string" , "value" : "nominal"},
  	{"position" : "4", "variable" : "fnlwgt", "storage" : "integer" , "value" : "continuous"},
  	{"position" : "5", "variable" : "education", "storage" : "string" , "value" : "nominal"},
  	{"position" : "6", "variable" : "education_num", "storage" : "integer" , "value" : "ordinal"},
  	{"position" : "7", "variable" : "marital_status", "storage" : "string" , "value" : "nominal"},
  	{"position" : "8", "variable" : "occupation", "storage" : "string" , "value" : "nominal"},
  	{"position" : "9", "variable" : "relationship", "storage" : "string" , "value" : "nominal"},
  	{"position" : "10", "variable" : "race", "storage" : "string" , "value" : "nominal"},
  	{"position" : "11", "variable" : "sex", "storage" : "string" , "value" : "nominal"},
  	{"position" : "12", "variable" : "capital_gain", "storage" : "integer" , "value" : "continuous"},
  	{"position" : "13", "variable" : "capital_loss", "storage" : "integer" , "value" : "continuous"},
  	{"position" : "14", "variable" : "hours_per_week", "storage" : "number" , "value" : "continuous"},
  	{"position" : "15", "variable" : "native_country", "storage" : "string" , "value" : "nominal"},
  	{"position" : "16", "variable" : "class", "storage" : "integer" , "value" : "nominal"}
  ]
}

Make sure the datasetID (here the value 9999999) is correct. To get the list of valid identifier, you can run Step 6: List all registered datasets from the Test the Data Set SAP Predictive services using a REST client
tutorial

Click on Send

The top 2 influencers are still the marital_status and capital_gain.

But if you look closer, you will notice that the capital_gain is now properly encoded as a continuous attribute.

However, you can see that one of the group for the capital_gain variable has a higherBound equal to 99999 which is in fact the missing value indicator.

Let’s try with the following Body, which define the missing value for the relevant attributes:

{
  "datasetID": 9999999,
  "targetColumn": "class",
  "numberOfInfluencers" : 2,
  "targetKey" : 1,
  "skippedVariables" : ["id", "sex", "race"],
  "variableDescription" : [
    {"position" : "1", "variable" : "id", "storage" : "integer" , "value" : "nominal" ,  "key" : "1"},
  	{"position" : "2", "variable" : "age", "storage" : "integer" , "value" : "continuous"},
  	{"position" : "3", "variable" : "workclass", "storage" : "string" , "value" : "nominal" ,  "missing" : "?"},
  	{"position" : "4", "variable" : "fnlwgt", "storage" : "integer" , "value" : "continuous"},
  	{"position" : "5", "variable" : "education", "storage" : "string" , "value" : "nominal"},
  	{"position" : "6", "variable" : "education_num", "storage" : "integer" , "value" : "ordinal"},
  	{"position" : "7", "variable" : "marital_status", "storage" : "string" , "value" : "nominal"},
  	{"position" : "8", "variable" : "occupation", "storage" : "string" , "value" : "nominal" ,  "missing" : "?"},
  	{"position" : "9", "variable" : "relationship", "storage" : "string" , "value" : "nominal"},
  	{"position" : "10", "variable" : "race", "storage" : "string" , "value" : "nominal"},
  	{"position" : "11", "variable" : "sex", "storage" : "string" , "value" : "nominal"},
  	{"position" : "12", "variable" : "capital_gain", "storage" : "integer" , "value" : "continuous" ,  "missing" : "99999"},
  	{"position" : "13", "variable" : "capital_loss", "storage" : "integer" , "value" : "continuous"},
  	{"position" : "14", "variable" : "hours_per_week", "storage" : "number" , "value" : "continuous"},
  	{"position" : "15", "variable" : "native_country", "storage" : "string" , "value" : "nominal" ,  "missing" : "?"},
  	{"position" : "16", "variable" : "class", "storage" : "integer" , "value" : "nominal"}
  ]
}

Make sure the datasetID (here the value 9999999) is correct. To get the list of valid identifier, you can run Step 6: List all registered datasets from the Test the Data Set SAP Predictive services using a REST client
tutorial

The performances are roughly the same as you limited the analysis to only 2 attributes on a small dataset.

Off course you can increase the number of influencers and see the difference.