As a test, I calculated the RMSE using simply (user average + item average)/2, on the same train test split provided in the code. For the Douban dataset I got RMSE=0.776, for the flixster dataset I got RMSE=0.958. Both better than the results in the paper.
Just wondering did you do a similar test.

Hi, this is a very impressive work. I just have one question, regarding the timestep t in in figure 4 for the evolution, what's the t specifically represent? Is that the number of iteration loop, or the diffusion step number in one loop?

I know you set T=10 for diffusion, but I assume the t in figure 4 is not the same with the diffusion timestep, otherwise it means you can improve RMSE from 2.26 to 0.01 in just one iteration of the nn?

Thank you!

How did you generate Wrow and Wcol that are contained in mat file ?

Can you please explain, how you deal with files (like u.data , u.item etc) to get the information about occupation of the user so, that model can recommend the movies related to the same occupation user interest?
Can you please explain , How you make split_1.mat file for movies100k dataset ?

Hi,

Is it possible to get the code you used to produce Figure 6 and 7 in the "Geometric Matrix Completion with Recurrent Multi-Graph Neural Networks" paper?

Cheers,