I am trying to reimplement SVD from the Netflix by using the movies rating.csv, meanwhile I have seen that you have done that on github https://github.com/WindQAQ/tf-recsys/ which is a great work. I used your package, and the RMSE is roughly 0.9 .
However When I am doing that by myself, I can't really get that level: my project is here: https://github.com/yangfeiran/svd
I can hardly push the RMSE below 0.95
I guess you are indeed good at tensorflow and SVD, so I hope it would be fine if I send you a email directly to ask my question :"where is wrong about my project?", cause the only difference I can see is that you are using keras from tensorflow, while I am not.
The reason I would prefer pure tensorflow instead of keras is that: in this way, I can make the code clear enough so everyone can see steps as tensor flows.
And I have created this question on stackoverflow, but no one answered it yet. https://stackoverflow.com/questions/50269129/svd-recommendation-collaborate-filtering-tensorflow
Looking forward for your reply,
Feiran (This is my facebook: https://www.facebook.com/bosilyang, I would be appreciated if you can add me on that)