I have data currently available that is very accurate and I would like to train my classification methods on this set of clean data to learn the important markers for distinguishing between classes. But in the future, my trained classifiers will not be seeing and performing decisions on this cleaned data; instead, it will likely have a lot more noise following some unknown distribution(s). Thus I am wondering, is it 'better' to train on noisy data if I'm going to likely see noisy data in the future, or train on good data since the noisy data should (ideally) correspond to the cleaned data if noise was removed?

Intuitively, if my goal is to simply perform classifications, then training on noisy data seems the 'better' approach since this is better representative of my expected future inputs. But if my goal is to learn about the data and what constitutes a particular decision, then training on cleaned data appears the 'better' approach.

But am I overlooking anything? Would training on the clean and/or noisy data be preferable for different reasons (e.g. generalization, overfitting, reducing dimensionality)?

  • $\begingroup$ Training on the clean data could definitely result in overfitting. If you don't have to choose, why not try both? Training on the clean data could help you understand the data you already have better. I would definitely recommend training on noisy data for predictions. Unless there's a way to clean incoming noisy data? What's the nature of the noise? $\endgroup$ Sep 12, 2018 at 20:55

1 Answer 1


The answer to this severely depends on what you mean by 'noisy' data. Are the labels noisy i.e. wrong? Or are the features noisy? Or both? If only the features are noisy, definitely use the noisy data and probably also the clean data. If only the labels, definitely do not use the noisy data. If both, is it possible to correct the labels? At the very least, can you get a reliable test set (representative with correct labels)? You could try training on both the noisy and clean data and see which gives better performance. Important methods to consider here would be regularization and early stopping.

It also depends on what algorithm you are using. Linear regression, for example, is unlikely to over-fit, while neural networks are extremely sensitive to noise, and there are a variety of approaches in between those two extremes.


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.