Support Vector Machine Question

[knimeoutjie]

Dear KNIME Cummunity,

I have previously used a RNN to model the following problem and can easily get above a 98% accuracy.

• Water runs into a cylinder at a constant rate
• Choose 10000 random sized cylinders: (variables r, l and t)random radius, length and duration (say seconds of constant rate (flow))
• I added another feature (vector sequence) of the progression of the volume in the cylinder over time
• My RNN could predict with high accuracy which cylinders would be full and which not after the random time t (in seconds)

Now I want to use a SVM to do the same job:

• Firstly, I did not use a sequence vector as input as in the case for the RNN; I did this because I don’t think SVM supports that kind of input
• For that reason, I kept the whole sequence as individual feature (variable)
• You can find my workflow in the public domain or attached KNIME_predict cylinder(9) SVM.knwf (2.5 MB)

knime://My-KNIME-Hub/Users/knimeoutjie/Public/Cylinders/KNIME_predict%20cylinder(9)%20SVM

When I implement the SVM as shown below, the hyperplane in the gutter is on one of the axis, which clearly means that:

• Either there is no solution or
• I am not implementing it correctly

Would appreciate any help and comments.

Many thanks,

Leon

[beginner]

I can’t really help just some thoughts:

• RNN’s are specifically made for time-series like problems
• Neural networks in general after having been trained are in essence just very, very complex mathematical functions.
• Neural networks can do multi-target predictions

This to explain why it works well with RNNs. You are simply predicting a relatively simple mathematical formula.

Since SVM can’t do multi-target, if you want to predict which cylinder is full I suspect you will need as many SVM models as you have cylinders as you can only give output for 1 target like full/not full or the actual volume contained in it.

[knimeoutjie]

Hi Beginner,

Thanks for trying to help.

The problem is framed in a way that the outcome is binary. Either the cylinder is ‘already full’ or ‘not full yet’. Each training sample has one of the two outcomes. For that reason I should be able to model it at least to some degree.

Anyone else has some input?

Thanks,
Leon