I use Oryx 1 for ALS train.
There are some generated users or item latent features with all zero (0.0) values.
It's very low percentage though.
In what cases will ALS algorithm drops zero latent features?
Also, we noticed that there are some users/items NOT in generation /input folder, but, they appear in the X or Y latent matrix.
Why this happens? Any thought ?
I think this could happen if a user only had 0s for input values, which in turn could happen easily if a user had only one data point that was a 0. Generally you would not find a user or item in the model that was not in the input. But was the user maybe in previous input or the previous model? I think there are some corner cases where this can legitimately happen but it should be rare.