I am trying to train a NN model in MATLAB to predict the amount of overflow for flooded junctions in an urban runoff system and I have 45 samples and 15 features. The issue is, I don't think 45 samples are quite enough for predictive modeling. Is there anything I can do to have a have good model without the need to augment my dataset (maybe training another model instead of NN is a better choice?), or should I just jump right into data augmentation techniques? Is data augmentation even a wise choice?


1 Answer 1


As I wrote yesterday, there are issues with data augmentation. If you have a small sample size, you put yourself at similar risk of overfitting as you would be fitting a complex model, and if you have a large sample size where that is not such a concern, then I question the need to synthesize artificial data.

With just $45$ observations, you lack the sample size to do sophisticated modeling like neural networks (unless you just want to learn the mechanics of writing neural network code). Unless this overflow follows a simple pattern, consistently strong performance is unlikely.

My suggestion is to work with a simple model like a linear regression on a few features, perhaps three to five features, following a rule-of-thumb for using one feature per $10$-$15$ observations. This is unlikely to achieve the kind of performance that you would get from sophisticated modeling on a large data set, but this is probably all your data will allow.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.