Aaron Meyer

Thanks @MarieRoald! @cyrillustan, I believe we are using the [N-PLS construction](https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/10.1002/%28SICI%291099-128X%28199601%2910%3A1%3C47%3A%3AAID-CEM400%3E3.0.CO%3B2-C), right? We are also generalizing this to work for >3-mode X. We will add an explanation to the docstring.

@JeanKossaifi @MarieRoald this should be ready for your review again!

@JeanKossaifi, yes, I saw that error. It looks like it is in the COVID example, which this doesn't touch, which is a little strange. Regardless, I'm looking into it.

Hi @wanglu2014—Do you have an idea of where you hit the memory issue? If it is during initialization, then try setting `init="random"` to avoid SVD initialization.

Indeed, this is a challenge with CP decomposition. This arises from the fact that ALS is dependent on a starting point, and so it becomes important to have a good...

It would be helpful to have some information about this in the docs somewhere. This comes up all the time.

Done for the base PARAFAC method. We could consider adding this for the other methods as well.

@huaiyizhao @kzhang2 this would indeed be a nice addition. We don't have specific plans at the moment. However, if you have specific tensor completion algorithms in mind, it could help...

Hi Polina—This is definitely something we've talked about, but we do not have a scikit-like interface for at the moment. You can do something like this by using the `fixed_modes`...

At least how I envision it, both `random_*()` and the random number generating object would be able to modify the result. For instance, if you want to change the random...