How to change the variable type and how to dichotomize in PA Expert Mode

[Former Member]

Hello everyone,

I am pretty new to SAP Predictive Analytics and I want to ask you some questions about it. Currently I am using the desktop installation of SAP Predictive Analytics 2.3.

My first question is if it is possible to change the type(nominal, ordinal, continous) of each variable in the expert mode manually? I found it on the Automated Analytics but not in the Expert Analytics. Unfortunately I wasnt able to find it in the sap expert user guide aswell.

Secondly, what do you think about dichotomizing for example a variable that contains a category or color according to the behaviour of the algorithms?

Will this may improve the results of the algorithms?

Thanks in advance and kind regards,

Christopher

[Former Member · Accepted solution]

Hi Christopher,

In SAP PA Expert mode, you don't have categories like ordinal or continuous or nominal. In Expert Mode there is textual, numerical, Date wise  data types.

The part dichotomizing, can you explain it in bit detail?

Regards

Ranajay

[Former Member · Accepted solution]

Hi Ranajay,

thank you for your quick reply!

According to your question. Imagine you sale different products in three colors. A possible structure of the data could look like this:

Product name	Product color
Product A	Green
Product A	Red
Product A	Yellow
Product B	Green

The question is now if the following structure of the data gain a better model with the different algorithms of SAP PA. 

Product name	green_flag	red_flag	yellow_flag
Product A	true	false	false
Product A	false	true	false
Product A	false	false	true
Product B	true	false	false

Do you know if there are some best practices for this kind of issue?

Kind regards,

Christopher

[Former Member · Accepted solution]

What exactly you want to predict here? Depending on that model will be selected.

[Former Member · Accepted solution]

That was just a simple example to show the dichotomous variable.

Maybe I want to predict if the sales figures of a new product will be above average in the product category.  

[Former Member · Accepted solution]

Then you need to have past historical sales data.. The data you are currently having is basically classification based.

[Former Member · Accepted solution]

Yes of course. I do know that. My point is just the best structure of the variable for the algorithms.

[Former Member · Accepted solution]

There is nothing like best structure, it all depends on your data. If your  historical data is just time series based and there is no other independent column affecting your target then you should be using time series based forecasting algorithm.

If you have multiple independent columns which might be affecting your target that is the sales, then you need to use regression based forecasting algorithm.

Like in the above comment if your data is having something like product color, you might wanna use regression based forecast.

[achab · Accepted solution]

In addition to Ranajay's answer:

in Expert Analytics, when using Automated Analytics "nodes" a data type mapping is made behind the scenes between the Expert Analytics data types in the document and the Automated Analytics expected data types. This is based on a data type analysis heuristic.

My 2 cents here

Best regards

Antoine

[Former Member]

Hi Christopher,

Your product color dimension consists of nominal categorical values which means that the values can not be ordered in logical way. Thus, you should transform it to dummy variables as you defined in your previous post, so your approach seems right.

Expert Analytics mode (Prepare mode) allows you to define your own calculated dimensions. Why don't you write an simple if statement for each color? Then you can train your model based on these dimesions...

Regards,

Eser

[Former Member]

Hi Eser,

Thank your very much for your reply!

I will try that.

Kind regards,

Christopher