Problem: Backward Feature Elimination Filter


I (think that I) have a problem with the Backward Feature Elemination meta node.

Both input ports are connected to the same source node. Filter and loop have the same columns. I've also set up the target and prediction column in the end node.

The Feature Elemination is running about 1h now and since I started the execution I get this kind of warn messages:

WARN      Backward Feature Elimination Filter     No features selected yet

What am I doing wrong?

Knime Version is 2.3.4, Windows 32Bit

Thanks for your help!

Have you configured the Backward Feature Elimination End node.

You can only configure the filter node after the loop has completed. You can ignore these messages or temporarily disconnect the filter node from the loop end.

I'm using a K Nearest Neighbor and I have about 270 different columns (type is double) How long does it normally take to finish the loop? Maybe I'm just impatient.


Yes I have configured the Backward Feature Elimination End node with my class column and the prediction column.


The problem is, that it seems that my loop is never ending. Also, I'm not able to disconnect or remove the connection resp. connection line between filter node and loop end.

If I execute only the loop (all nodes on status yellow, with left clicking on Backward Feature Elimination End node and pressing F7), I still get the same warn message (filter node is on status red, but nevertheless it seems to be executed with every iteration).

Now I get also this warn message:

WARN      Backward Feature Elimination Filter     No features selected yet
WARN      TableContentModel     Exception while setting new table.


As Thor said the Backward Feature Filter node is only configurable once the Backward Feature Elimination Loop is complete, as it needs to know the results from this. Hence the warning message, this always happens.

When I have used this on around 25 columns it takes about 10 minutes, so I would expect 270 columns to take considerably longer.

Just trying leaving it, and eventually the loop should complete and the filter node goes amber and thus can be configured.


According to the docs, the number of iterations is (n - 1) * (n - 2) / 2 , which is ~36000 for n=270. If you time one execution of your loop body and multiple it by this number, you should have some estimate. Also note that v2.3.x has very slow GUI updates in loops (the loop execution is often much faster than what the GUI tells you ... something that is fixed in v2.4 -- the same is true for "Exception while setting new table").

hm strange..normally the number of iterations would be the gaussian sum +1 ?


for 10 attributes:

1 (starts with all) +9+8+7+6+5+4+3+2+1 = 1+      (Z * (Z-1) / 2)

here Z is the number of features (columns).

so for 10 attributes there are 46 iterations

with the formula in the description it would be

(10-1) (10-2) /2 = 36  (which is wrong!? isn't it?)

I wondered about that earlier. Maybe the description is wrong. I thought the alg leaves on more column out in each iteration?