Sorry if this might sound as a super stupid question but I’m a novice with KNIME.
I am trying to use the logistic regression learner node to analyse some of the data I have. It works perfectly fine when I’m using it the first 2 times, and then, on the 3rd time, when I try to exclude just 1 variable, it stops working and won’t execute, showing in the console the following error message: “ERROR Logistic Regression Learner 2:48 Execute failed: The logistic regression model cannot be computed. See section “Potential Errors and Error Handling” in the node description for possible error causes and fixes”
I can’t for the sake of me understand why, as the configuration is the same as in the 2 previous ones?
Is it something with the variable I’m trying to exclude?
I’m attaching screenshots of my current workflow + the configuration of the nodes:
The “Iteratively reweighted least squares” algorithm works well for many datasets but for some it struggles due to numerical problems (things become non-finite due to numerical under/overflow) and it seems like your dataset is one of these.
In order to find the actual cause I would need to play with the data but from the looks of features one problem might be that features might be highly correlated which can mess with the learning algorithm.
Could you calculate the feature correlations and remove columns that are strongly correlated (or rather keep only one of them)?
The “Stochastic average gradient” algorithm is more robust to this kind of issues because it takes smaller steps towards the solution and it also allows to use regularization.
In your case I would try to switch solvers but make sure that you normalize your data properly before you feed it to the learner node.
To do this simply use the normalizer node with z-score normalization.
I would also add some regularization e.g. a Gauss prior with variance 0.1 which makes sure that your coefficients can’t grow arbitrarily large.
In fact, I found out that the features where highly correlated, hence why it wasn’t working then.
I just couldn’t figure out before your comment.
For the goal of my observations, this result is more than enough tbh, but I’m going to implement your suggestions as I’m really curious to investigate further on this dataset.
