After reading the following CrossValidated post, I cannot derived the correct multiplicative rules for regularized NMF from this paper. They obtain the coefficients $|I_u|$ and $|U_i|$ in the denomiators of Eq. 14 (or below in red). To my understanding, the derivative of the $L2$ regularizers should contain no sum and thus the results should be the following (without the coefficients in red):

$p_{u,k} \leftarrow p_{u,k} \frac{\sum_{i \in I_u} q_{k,i} r_{u,i}}{\color{red}{|I_u|} \lambda_P p_{u,k} + \sum_{i \in I_u} q_{k,i} \hat{r}_{u,i}}$


$q_{k,i} \leftarrow q_{k,i} \frac{\sum_{u \in U_i} p_{u,k} r_{u,i}}{\color{red}{|U_i|}\lambda_Q q_{k,i} + \sum_{u \in U_i} p_{u,k} \hat{r}_{u,i}}$.

I guess the equations with the red coefficients is more or less the same except for a rescaling of the regularization weights. However, I was wondering why they decided to go this way.


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.