DL4J poor results

[alfroc]

Italiano

Hi,
I would like to use the DL4J Feedforward Learner node to achieve better (or at least similar) results than the RPROP node.
I used for that the well-known German Credit Dataset. After many attempts, I have not been able to achieve any decent result using that node.
Could someone tell me what I’m doing wrong?
Thanks!
AlfredoRPROP & DL4J.knwf (227.2 KB)

[ScottF]

Hi @alfroc -

I would suggest setting up your network similar to the one shown in this example workflow:

Example of a simple Multi Layer Perceptron – DaveK

In particular, you need many more units in your hidden layers, and your learning rate should be lower. You could play around with different activation functions too, although the example uses pretty much all ReLU. Also, set your number of training iterations in the learner node higher as well - maybe 20 or so.

I was able to get results marginally better than the RProp MLP nodes with a few changes - 82% vs 78% AUC - using settings like the above.

[alfroc]

Thank you, Scott, in fact I got better results!