I have learnt that handling matrices is not that simple in KNIME.
• I have the below sequence(s) to vector mapping. (for input in a RNN)
• It contains only two samples, but some things can be learnt from it
• The sequences have been padded to be the same length
• The output is different for the two samples
• The first and second parameter changes much less than the third, but I need all three since they are sequential
• This is a time series
How can I send this matrix to the Keras Network Learner? I specifically don’t want to make three input layers, but it would be interesting to hear your comments on the mathematical differences between the two (3 vector versus 1 matrix).
I believe image processing can help; is that true?
before I dive into handling matrices, I’d like to point out that you should consider normalizing your data, so that all your features are within the same (small range). That’s because a neural network is scale-sensitive. As a consequence the feature with the largest scale is likely to be the most influential even though it might actually be not very informative for your task.
Now regarding matrices in KNIME:
I still believe that the cleanest solution is to use 3 separate input layers which are then concatenated, mainly because you can name the input layers accordingly, which can help to prevent any potential issues once you carry over the network to deployment.
Anyway, I was wrong about the necessity KNIME Image Processing. You can still use the DataRow to Image node but in the end there is no difference to simply putting all your features into a single large collection and feeding it to the learner.
Note that it needs to be a flat collection i.e. not a collection of collections, otherwise the learner will complain.
