How to apply PCA to training/val/test data

Hello. I am working on a neural network that predicts anxiety using heart rate and heart rate variation (supervised machine learning). I have 232 samples and a total of 17 features. In the future, I want to be able to add samples and use this NN to predict anxiety for those. Obviously that is too many features, so I want to do feature reduction using PCA, but as much as I read, I am still a little confused.

I applied PCA to the dataset and 94% of the variation is in 8 components, but that was for the whole dataset. I know that I need to calculate PCA for the training and then do a thing for the validation and test data but I am so confused as to what that thing is and how to do it. Then redo that thing for new samples in the future. If someone could spoonfeed me an answer I would be so happy.

[Greg Heath]

1. Nomenclature: Delete the word cross.

validation and cross-validation are two separate animals.

Greg

[Nis Jar]

Ok. What then?

[Simon Schmid]

I think you need to use the same PCA model, which determined with your training data on the test and cross validation data.

Lets say you used following code:

[pcs,scrs,~,~,explained,mu] = pca(data);

%here you can use your first 8 components in scrs (with regard to the %explained variance in explained)

%Then use the model for your training and cross val. data (datatrain)
%and take here again the first 8 components

datatrain1=(datatrain-mu)/pcs';

% Can anybody confirm this code??