DL4J 特征重要性

Question

I have built a LSTM neural network with multiple features with Deeplearning4j. Now I want to know if a feature which I have added is good or not. I remember from other programming languages that there is a possibility to get a percentage value (feature-importance) per feature.

How to get this list of features from a model?

Answer 1

Feature Importance isn't commonly used with neural networks in general regardless of framework. Generally the focus should be on hyper parameters and viewing the neural network training in the UI to see how its gradient changes over time.

From there, you see if the loss curve has something like a:

| |
| |
| |
| |
| \
|  \________________________________
------------------------------------------

loss curve. You won't always get that but anything where the training is smooth and stable indicates well tuned parameters, the right loss function being picked and the learning happening.

The only other thing I can recommend is ensure that your input data as needed is properly normalized.