I perform matrix factorizaition in my data using the sklearn implementation of Non Negative Matrix Factorization. In the evaluation process I am removing some values from my initial dataset and I am trying to see if the approximation matrix of NMF can predict well those missing values. What I have notice from the experimental results is that, the smaller the latent factors of the decomposition are the better the prediction of the missing value I got. However, this means also that the rest of the values are not approximated really well. Is this an expected behavior? Why this is happening? Moreover, is there any explanation for the l1_ratio and the alpha value?

  • $\begingroup$ Can you please clarify with more detail the validation scheme you are using? It might just be the case you are looking at an inappropriate metric (happens). $\endgroup$
    – usεr11852
    Mar 21 '17 at 21:47
  • $\begingroup$ an answer to @usεr11852 would be needed. It sounds like overfitting of the model, which would give exactly the behaviour you describe and understanding your validation scheme would help diagnose the risk and perhaps allow suggestions on how to fix it. $\endgroup$
    – ReneBt
    Mar 16 '18 at 9:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.