Impute Missing Values

For this operator "Impute Missing Values", you can treat it as an iterative loop. Inside the loop, for each iteration, a machine learning model (here KNN) will be trained to predict missing values in one of the columns of your data. You don't have to manually define the label with "set role" because it is auto configured.

Decision Tree will be configured to predicting nominal label (GAIN_RATIO INFORMATION_GAIN, GINI_INDEX_ACCURACY), or numerical label (LEAST_SQUARE) according to the criterion parameter. You can not use one Decision Tree model (fixed criterion) to impute both numerical and nominal columns inside this nested operator. However, K-NN is different story, which is powerful to handle both numerical and nominal labels. Other candidate models to consider are GLM, GBT. 

Check out the operator info here

As explained in the first reply, the learner is built iteratively on each column. When you impute for column A, it automatically set column A as label because you need to predict the missing values in column A. When you impute column B in the next round, the learner will use non-missing values of column B as label to predict the missing values in column B. Repeat this (set different column as label in each step) for column C, column D, column E,…, until you finish imputing missing values in all columns.
More explanation and implementation details can be found on the GitHub open source page here
https://github.com/rapidminer/rapidminer-studio/blob/master/src/main/java/com/rapidminer/operator/preprocessing/filter/MissingValueImputation.java

What is role? Check this out https://community.rapidminer.com/discussion/54288/roles-and-labels-a-quick-guide

You should insert break points before the learner inside the nest and check the refreshed metadata in each iteration. In your sceeenshot, the metadata is valid for the first iteration Only.

Thank you for your comprehensive replies, including directing me to the operator code on the GitHub, which clarifies a lot, particularly "* setting one of the regular attributes to label under the assumption that all * attributes are from the same type".
I think we could set the question as solved from practical point of view (although it could be interesting to investigate the case when the above assumption is not hold while a learner accepts attributes of a specific type, like DT for a selected criterion; maybe at least an explanation in the IMV operator description in Help would be helpful, if not enabling the IMV to impute missing values for attributes of that specific type)