Is it a common practice to do clustering before supervised learning to eliminate "noisy data"? Obviously, depending on the type of task. It seems like it makes sense in my case and my neural network then achieve better results.

Is there any resource (paper, online article, or anything else) regarding this?

  • 1
    $\begingroup$ In general, data whether 'noisy' or not, should not be removed unless it represents a clear error. If you eliminate data that is difficult to explain, you will artificially inflate your metrics of performance. $\endgroup$
    – mkt
    Aug 4, 2022 at 10:33
  • $\begingroup$ I know, but if only removed from the training set they should not impact performance in that sense $\endgroup$
    – mlp
    Aug 5, 2022 at 12:35
  • 1
    $\begingroup$ They will very likely reduce performance on the test set if removed from the training set. Removing from both training and testing set will likely leave estimated performance unaffected, though that estimate will be biased. $\endgroup$
    – mkt
    Aug 5, 2022 at 14:38