In Rapid Miner Go the linear regression algorithm used some inputs I did not select.

BillP

I hope my question has not been asked before. In short, RapidMiner Go seems to be running a regression with variables I did not select. An explanation follows. In Rapid Miner Go I dropped a csv files with 64 columns and almost 2900 rows. I wanted to predict a single column (of numbers) using linear regression and decision tree ("Easily Interpretable"). The first two columns were date and time. The other columns were numbers. I selected only 5 inputs and an indicator on that page said 5 were selected. I ran the regression and in the Data Metrics it reported the correlation for the 5 inputs that I selected plus 7 others I did not select. Assuming that it ran the regression with the 7 inputs I did not select how do I run the regression with only the 5 inputs I selected? Thanks very much. Regards, Bill

varunm1

Hello @BillP

Can you cross-check if the model is built on more than what you selected? You can do so by clicking on the model link after it executes and then scroll down to see how many attributes are there with coefficients.



Coefficients checking:

BillP

Thank you Varun. I did what you suggested (actually I have checked the coefficients many many time) and there are 12 coefficients not 5 as there should be. The 7 extra inputs seem to have been randomly chosen. When I go to Model Simulator and move the slider corresponding to the inputs that I did not select over their entire ranges, they will move the predicted variable by a small amount, 0.01 or 0.001%. The "weight" for some of them seems significant but moving the slider for that variable doesn't move the predicted value very much. It is as if the regression went haywire. I don't know why.

BillP

I suppose it doesn't matter, but the fit is very good. The plot of model versus actual value looks a lot better than I expected it would be.

varunm1

That seems weird. Can you do the following and provide me your process?

In Rapidminer Go, once you click on the model link as inform earlier. You have an option called "Export" on top right corner. If you click on that, you will have an option called "Download Process". Can you download that process file and attach it here to check?

BillP

Thanks. I attached the exported model.

varunm1

Hello @BillP

Thanks for sharing this. I will take a look, also if possible try to share your data here or in a private message so that I can rerun and explain to you the reason for his phenomenon.

BillP

Hi Varun, Thanks so much. By private message do you mean your private email at your web site?

varunm1

I got your email. In the future you can use the message option on the rapidminer community as well. If you click on my name it will take you to my profile, you can find "Message option on Right corner. You can send a message attaching the file. Sample image below.

BillP

Thanks!

BillP

As you indicated in your private message to me, I removed commas from the 6 column headings that I did not select but still ended up in the analysis. After removing the commas, the linear regression model results showed only 5 inputs and the estimation results were fantastic. I hope this info about commas in the column headings helps anybody that might have the same problem. However, I don't think anyone in the future will find this thread unless the subject is something like "Commas in column headings cause a problem when selecting inputs". I am a little surprised that this was not caught long ago. I use commas to separate a label and the units of that label such as "Mass flow, t/h". The following should not cause a problem in RapidMiner Go: "Mass flow [t/h]" Thanks so much for your help and have a good day!

BillP

How do I credit you Varun with the answer? I can't click yes on my comment because I just acted on your advice

varunm1

Hello Bill,

No problem, Lets keep this question open as I want our friends at RM to check this and maybe open a ticket to resolve this comma issue. I am not sure if there is already a NOTE that says we cannot use comma in attribute names but will wait to see this getting resolved so that there won't be future issues for anyone.

@sgenzer any inputs here?

BillP

If there are commas in the column headings should there be a warning to the user when they are detected. It is not that inconvenient to remove them, once you know that they should not be there. The inconvenience is finding out that information. Congrats on noticing the commas, that was very astute. 

varunm1

Hello @sgenzer ,

In order to repeat this error, please upload this CSV file to RapidMiner GO and select "angle" as prediction variable and attributes in the below image (which doesn't have comma) and use default selections in the next window, Easily Interpretable and everything left as default and run the analysis.



Once the analysis is done, we can observe the GLM model also used unselected attributes as shown below. 



The reason observed is related to the presence of "," comma in the attribute name. My understanding is that REGEX function present in the Load and Process Data --> Remove column module is being tricked by this comma value. I don't see this once the comma is removed from attribute names. Also, with a comma in the attribute name, this doesn't happen in the auto model.

I am not sure if there is an instruction to not use comma in attribute names.

sgenzer

thx @varunm1 I'm pushing this to Prod Feedback and will report to the RM Go team.