Make `NeuralNetworkClassifier` to accept categorical input

[adekusar-drl]

Make `NeuralNetworkClassifier` to accept categorical input

[mattwright99]

@adekusar-drl I am interesting in picking up this ticket. Here is what I am thinking right now:

• Add a method to NeuralNetworkClassifier that has a parameter categorical_cols which is a list of column indices that are then one-hot encoded. unless there a way you can think of to convert categorical data automatically?
• Call this method at the start of fit, predict, and score.

Alternatively, I could implement this function in the dataset_helper script or somewhere else so it can be used in other models or by the user as a preprocessing step.
What do you think?

[adekusar-drl]

Hi, in general you are right. So, first of all, please take a look at OneHotEncoder from sklearn, I hope we can make use of it instead of implementing a conversion login from scratch. Next, I'm not sure we need categorical_cols right away, we have a basic classifier that does not accept different type of labels, so should be easy to derive a label type from true labels.
Anyway, any suggestions are very welcome as some exploration is required here.

[mattwright99]

Oh I see, I thought you were refering to categorical features not labels. That should be easy enough to infer and transform.
I will let you know if I have any other questions!

[adekusar-drl]

I don't know why I was thinking of labels only. Categorical features can be also a case, but such feature would require even more exploration and investigation. There's also LabelEncoder in sklearn that may help.

[mattwright99]

Ya categorical features would be interesting but slightly more complicated. I think we would either need to have the user preprocess the data before instantiating the model to specify number of qubits or we would have to use PCA to force the encoding dimensionality to fit the number of qubits. Could be a cool task but might warrant its own issue/PR.
And yes I have been looking into sklearn's LabelEncoder and OneHotEncoder. Thanks for the recommendation.

[adekusar-drl]

On the top level, I think, we should keep interfaces as simple as possible and, by default, users should not bother of number of qubits and so on. And in the same time the interfaces should be flexible, so experienced users can tweak model in a way they want. Yeah, feel free to split the work into two or more PRs.

[mattwright99]

@adekusar-drl while we wrap this ticket up, I would be interested in exploring how to automatically handle categorical features, if you think that would be valuable. Otherwise I would be happy finding another issue to take a look at.

Thoughts?

[adekusar-drl]

Certainly, you can explore categorical features, but, honestly, I don't know what you may hit on this road.