Hi, Stolee,
I use your code to reproduce the NNMF. However, I can't get the same experimental results as your paper "Matrix Factorization with Neural Networks and Stochastic Variational Inference". The test RMSE
(ML-1M) of NNMF is 0.8539, my experimental result is 0.8725 with the parameters (lambda = 1 (Obtained by hyper-parameter selection), learning rate = 0.005, batch-size = 25000). Can you tell me the parameters(RMSE 0.8539 on the ML-1M)?
Hi. First of all, I am very impressed of your implementation of NNMF model.
You said, in the SVINNMF paper, that the RMSE result of NNMF model on ML-100K is 0.9380. Right?
But, in my experiments, I achieve the RMSE 0.906 on that data. This result is consistent with the original NNMF paper's claim.
I think that you should you RMSPropOptimizer with learning rate 1e-3 and full batch including bias on MLP(I think that you're aware of this issue seeing the fix branch on this repository). I think these will leads you to enhance performance of your NNMF implementation.
I already uploaded my implementation. If you interested in this issue, please check my commit d63e7b1 in my NNMF repository and feel free to contact me. Thanks ๐
p.s. I really appreciated of your efforts on implementing NNMF model.