I’m currently getting the following error message in the Logistic Regression Learner node:
the algorithm did not reach convergence after the specified number of epochs. Setting the epochs higher might result in a better model
I’ve gone to the advanced setting tab and increased maximal number of epochs to from 100 to currently 1,500. I’ve reduced the number of variable (now six to see if this would help) do you have any suggestions?
Thank you.
Ian
As a rule, I suspect that this is related to the dataset that you are using. A few notes to consider below:
- Remove/handle the features that may be highly correlated with one another
- Transform numeric variables (min/max, z-scale)
- Different solvers
- Increasing the number of epochs
To see convergence, you could always start very small. Your binary outcome variable against a single input feature. Iterate from there.
Does the Stochastic average gradient solver in Knime assume the independent variables in the logistic regression are normally distributed? Please see the file attachment Example Variable where this variable was put into bands and is clearly positively skewed.
The raw data has been normalisation in Knime before modelling.
Regards.
Ian
Sorry I should have added that most of the predictive variables in the dataset are likely to be positively skewed in this manner.
Thank you.
Ian
Hi there,
I’ve now taken a random sample of the dataset so now only using about 10% of the dataset and also stratified the dataset so the dependent variable is in a 1:3 ratio of event to non-event. I’m still getting an error message:
“the algorithm did not reach convergence after the specified number of epochs. Setting the epochs higher might result in a better model”
This morning, reweighted the dataset so the event and non-event are now 50/50 and still getting the above error message.
Regards.
Ian
Hi there,
I’ve now used the alternative logistic regression algorithm available in Knime, Iteratively reweighted least squares. This has worked and is performing well.
Regards.
Ian