I have a cople of simple questions. In the Scorer note description there is the following sentence " ..It is possible to hilight cells of the matrix which propagates highlighting to the corresponding rows. Therefore, it is possible for example to identify wrong predictions " .
In other words, if I understand correctly, It is possible not only to get the percentage of wrongly-predictied data (for instance in a decision tree model) but also to identify exacly those data.
Q1: Did I understand correctly?
Q2: If yes , how can I get these detailed data ? I tried to click on the confusion matrix, but probably this is not the way...
Thanks a lot!
In addition to the output ports there is an option in the context (right-click) menu for the node named "View: Confusion Matrix". This option will open up an interactive view which has the key feature of enabling highlighting. If you select a cell in the resulting view, you can use the hilite menu to demark these events and via magic KNIME will push those hilites as far back in your workflow as possible. If you just want to filter the hilited results simply include a hilite filter attached to the output of either your predictor or further upstream if you prefer.
Let us know how this works for you.
I've tried to implement your advices, but I'm not sure I've got everything.
1. I connected an Hlite Filter node to my Decision Tree Predictor nodes, after having highlighted some branches in the trees and the Hlite Filter seems to work properly. I use the expression seems as I'm not sure all the detailed data included in the leaf are visualized... But I'll check it carefully
2. I highlighted a cell in the confusion matrix (for instance the bad classified data, using the open views menu ) and connect the Scorer Node to an Hlite Filter Node, but nothing happens, that is , the node does not show any one of the highlighted records.
For sure I did something wrong...
If you want to filter records that have been highlighted, you will need to hook up the highlight filter to the table that contains the records of interest, not the scorer. That way when you filter you are filtering on the records and not the conusion matrix. Does that make sense?
The node layout in the post above shows you how to use a highlight filter with the reults of the scorer.